An off-line detection method for automotive acoustic system

Abstract

The present invention relates to an off-line detection method for an automobile acoustic system, comprising: S1, play acquisition, testing a hybrid system composed of a loudspeaker and a microphone, the system drives the loudspeakers one by one to play a set signal, synchronously controls the microphone acquisition, and obtains recording data; S2, transfer function calculation, calculating the transfer function between the loudspeaker and the microphone; S3, model training, model training is performed by an unsupervised learning method to obtain a Gaussian mixture model; S4, judgment, judgment is performed by the trained model. Since there is no need to deploy additional microphones and loudspeakers for separate testing, and the microphone and the loudspeaker can be tested at the same time, the test time can be saved, and it is suitable for testing the acoustic system in the car in a short time after the car is off the line.

Description

Technical Field

[0001] The invention belongs to the field of detection, and relates to vehicle detection, in particular to the testing of loudspeakers and microphones in automobiles. Background Art

[0002] In the automotive industry, the acoustic system is an integral part of the vehicle, which includes functions such as on-board audio, voice recognition, and noise control, providing drivers and passengers with a better driving experience. In order to ensure the reliability and performance of the automotive acoustic system, manufacturers need to implement effective end-of-line testing methods during the production process. End-of-line testing refers to the comprehensive testing and evaluation of the acoustic system after the automotive assembly process is completed to ensure that it meets the design specifications and quality standards.

[0003] The current off-line testing methods for automotive acoustic systems mainly include sound quality testing, frequency response testing, and voice recognition performance testing, etc. These include testing of the acoustic components themselves, as well as testing of the functions of the acoustic components. For example, the frequency response test is a test of the speaker and microphone themselves, while the voice recognition test is a test of the functions.

[0004] The electroacoustic performance tests for speakers and microphones mainly include frequency response tests, distortion tests, scratch sound tests, etc. Such tests usually require a relatively complete acoustic environment, such as an anechoic chamber. In the development stage, they are mainly conducted in the early R&D stage or at the component level.

[0005] At present, the number of speakers and microphones in cars is increasing. As of now, the number of speakers in high-end cars has increased from 6 to about 20. The installation locations are also different.

[0006] In a car, speakers are required to create a good surround sound effect for the whole car. Speakers with different functions are distributed in different positions. The same is true for microphones. Currently, due to the demand for voice control and the popularity of active noise control technology in the car, the number of sensor microphones and reference microphones is also increasing. They are also installed in different parts. This brings the following main problems:

[0007] (1) The interior of a car is not a standard acoustic environment, and it is difficult to test acoustic performance. The cutoff frequency inside the car is high, and there is almost no free field radius, so standard acoustic testing cannot be performed;

[0008] (2) A car is a complex mechanical component with serious tolerance accumulation, which may eventually affect the acoustic performance. Although each individual component including the speaker (microphone) is tested separately, the installation tolerance or stress can only be finalized after the whole assembly is completed, so the most appropriate time to conduct acoustic testing is when the product is finally offline;

[0009] (3) The shipment volume of new energy vehicles is gradually increasing, and the popularity of fully automated production lines means that after the final assembly is completed, there is only a short period of time for testing. During this time, the car is still on the conveyor belt, and personnel are setting up additional testing equipment. Whether in terms of safety or operational feasibility, the reliability of the test is poor.

[0010] Due to the above-mentioned restrictions, there are currently few acoustic tests on the final offline products, and the main requirement is to determine whether the installation is loose or causing abnormal noise.

[0011] At present, acoustic testing of cars after they roll off the assembly line is inconvenient to carry out, and with the increase in shipments, it mainly relies on subjective listening. This brings about greater test uncertainty. At present, there are attempts to use a method of laying out additional test equipment inside to conduct testing, but due to the different fixing methods of different cars, additional fixtures are required, and it is very challenging to arrange, test, and dismantle them in a short time after the whole assembly rolls off the assembly line. Summary of the invention

[0012] The purpose of the present invention is to provide an off-line detection method for an automobile acoustic system, which solves the problem of how to quickly perform an off-line test on an automobile.

[0013] In order to achieve the above object, the technical solution adopted by the present invention is:

[0014] The present invention provides an off-line detection method for an automotive acoustic system, comprising:

[0015] S1, play and collect, test the hybrid system composed of speakers and microphones, the system drives the speakers one by one to play the set signal, and synchronously controls the microphone to collect and obtain the recording data;

[0016] S2, transfer function calculation, calculating the transfer function between the speaker and the microphone;

[0017] S3, model training, model training is performed through unsupervised learning methods to obtain a Gaussian mixture model;

[0018] S4: Judgement: Make judgment through the trained model.

[0019] Preferably, the speaker in S1 generates a white noise signal.

[0020] Preferably, during excitation in S1, a low-pass filter is performed to filter out high-frequency signals above 1 / 2 of the sampling rate of the corresponding microphone setting. Since different microphones have different uses, if the corresponding sampling rates are set differently, they cannot be collected synchronously here, and microphones with the same sampling rate need to be grouped as one group, and microphones with different sampling rates as another group.

[0021] Preferably, in S2, a finite impulse response model estimation method is used to calculate the transfer function.

[0022] Preferably, in S2, feature extraction of the transfer function is performed through time domain features and frequency domain features, and then the extracted features are detected.

[0023] Furthermore, in S3, the amount of computation is reduced by reducing the dimension of features.

[0024] Preferably, during the model training and fitting process of S3, the real vehicle label of the vehicle is manually determined, or the label is obtained by listening to a recording.

[0025] Preferably, in the judgment of S4, the status of the vehicle off the production line is judged by feature extraction and then by the model fitted in S3.

[0026] Preferably, it also includes a model updating step, re-listening to the judged data or customer complaint data, re-confirming the labels, and then updating the training database, and finally updating the Gaussian mixture model.

[0027] Due to the application of the above technical solution, the present invention has the following advantages compared with the prior art:

[0028] The off-line detection method for the automobile acoustic system of the present invention does not require the deployment of additional microphones and speakers for separate testing, and can test the microphone and the speaker at the same time, thereby saving testing time. The method is suitable for testing the acoustic system in the automobile within a short time after the automobile is off the production line. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Hereinafter, some specific embodiments of the present invention will be described in detail in an exemplary and non-limiting manner with reference to the accompanying drawings. In the accompanying drawings:

[0030] Figure 1 This is a typical diagram of the in-car speakers, microphones, and body structure of a new energy vehicle;

[0031] Figure 2 A transfer function in FIR form from a single speaker to a single microphone.

[0032] Figure 3 Yes Figure 2 The system frequency response function is obtained by periodically extending the transfer function shown;

[0033] Figure 4 is the characteristic distribution map;

[0034] Figure 5 It is a flow chart of judgment;

[0035] Figure 6It is a process diagram of model updating and training. DETAILED DESCRIPTION

[0036] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. 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 are within the scope of protection of the present invention.

[0037] The method used in this example is to Figure 1 A typical new energy vehicle, the speaker, microphone, and chassis are as follows Figure 1 As shown in the figure, the microphone and speaker are both controlled by the vehicle computer.

[0038] The implementation is divided into the following steps, which are mainly divided into two stages: training stage and prediction stage. The training stage is mainly divided into the following steps:

[0039] 1. The car computer drives the microphone to collect data and drives the speakers to play specific signals one by one;

[0040] 2. Calculate the transfer function of each speaker and microphone;

[0041] 3. Data is uploaded to the server;

[0042] 4. Model training is performed in the server using unsupervised learning methods.

[0043] The prediction phase is similar to the training phase and is divided into the following steps:

[0044] 1. The car computer drives the microphone to collect data and drives the speakers to play specific signals one by one;

[0045] 2. Calculate the transfer function of each speaker and microphone;

[0046] 3. Make judgments based on the trained model.

[0047] In addition to the data upload part, it is mainly divided into the following three aspects:

[0048] 1. Playback collection;

[0049] 2. Transfer function calculation;

[0050] 3. Model fitting / judgment.

[0051] Capture and playback

[0052] Playback and acquisition are essentially the process of testing the transfer function of a hybrid system composed of a speaker and a microphone. In the actual working process, the linear performance of the device is mainly examined. If the transfer function of the speaker is S(s) and the transfer function of the microphone is M(s), then the ideal system transfer function is H(s) = S(s)M(s), but usually, the system will introduce other transfer functions D(s) through the operating system and other electrical processes. Usually, D(s) mainly manifests itself as a delay link. Usually, system testing is mainly to test the transfer function of the speaker or microphone alone (more mainly to evaluate the frequency response function, etc.), and the transfer function is equivalent to the unit impulse response function of the system, which can be obtained by Fourier transforming the unit impulse response function. Note that the speaker here preferably emits a white noise signal, and because the playback frequency of the speaker inside the car is usually 44.1KHz / 48KHz, and the microphones have different divisions of labor, 8000Hz, 16000Hz, and 48000Hz are all used. Therefore, in order to avoid frequency aliasing, it is necessary to optimize the playback frequency that each car microphone can use during excitation, and perform low-pass filtering to filter out high-frequency signals. Taking the microphone acquisition frequency of 8000Hz as an example, it is necessary to filter out high-frequency signals above 4000Hz and 1 / 2 sampling rate.

[0053] The general process can be expressed as:

[0054] 1. The system queries the sampling rate supported by the microphone. If the microphone driver supports the same sampling rate, it will collect data at the highest sampling rate. If not, it will collect data from the microphone alone. The microphone sampling rate is recorded as f m1 , f m2 ,….

[0055] 2. The system generates a random seed noise signal and filters it using 1 / 2 the microphone sampling frequency.

[0056] 3. The system drives the speaker to make sound and synchronously controls the microphone to collect data (usually two instructions in the program, there is a certain delay, but it can be ignored), and obtains the recording data in the form of s1m1.wav (the wav file format is not required, and an array relationship that represents this relationship can be used). The processed excitation signal here needs to correspond to the microphone.

[0057] 4. Determine whether there is another pair of speakers / microphones and repeat step 3.

[0058] Transfer function calculation

[0059] The transfer function calculation here essentially belongs to the category of system identification. In this application, we use the finite impulse response model estimation method. The system to be identified is For the system to be identified, the FIR model is used to describe the response of the system:

[0060]

[0061] Where u(k) is the input of the system, g(l) is the filter coefficient. When the signal used to calculate the transfer function is large enough and the signal to be identified is stable enough, the FIR model can be well approximated to the original model. When calculating the FIR model, it should be noted that due to this FIR function, g(l) is actually equivalent to the unit impulse response function of the system. Its frequency response can be obtained by performing fast Fourier analysis on the above g(l). Therefore, the order of g(l) needs to consider the frequency resolution of the FFT analysis, and it is better to be a power of two.

[0062] The tuning process of g(l) can be carried out using the LMS algorithm. The following formula is a commonly used LMS coefficient update method, where μ is the learning factor and g t is the filter coefficient at the corresponding moment, u(n) is the input at the current moment, and e(n) is the error at the current moment:

[0063] g t+1 =g t +2μu(n)e(n)

[0064] The execution flow of the above algorithm is as follows:

[0065] 1) Filter initialization: g, u (filter array and input array)

[0066] 2) For each new input sample u, calculate the output signal

[0067] 3) Using the expected output, calculate the error signal e(n) and get the gradient

[0068] 4) Update the filter coefficients using the LMS update formula: g

[0069] 5) Return to step 2) until the end, and the output sequence and error sequence can be obtained.

[0070] Model fitting / judgment

[0071] Model Fitting

[0072] After the above two steps, we have obtained the FIR transfer function from a single speaker to a single microphone, which usually has the following characteristics: Figure 2 As shown. Extending the cycle can obtain Figure 3 The frequency response function of the system.

[0073] They correspond to the time domain response and frequency domain steady-state response of the system. Some features are established to detect the transfer function.

[0074] With the rise of machine learning and artificial intelligence, probability theory is a powerful tool that enables handling uncertainty in many applications from classification to prediction tasks. For scenarios such as speaker microphones, supervised learning in automotive scenarios requires the establishment of a large database, and due to the rapid iteration of vehicle models and large differences between models, supervised learning is not a good method. Unsupervised learning automatic clustering is more suitable for this scenario. The Gaussian mixture model can approximate the probability distribution of any shape and has good applicability. The distribution of each feature can be roughly as follows Figure 4 to describe the distribution of .

[0075] In the final offline process, it is usually not necessary to perform multi-classification on the results. Binary classification is sufficient, that is, OK or NG. In this way, the corresponding features of OK should have distribution characteristics and cluster within a range, while the number of NG is relatively small. By fitting a Gaussian model to normal samples, the discriminant model can be obtained.

[0076] Usually Gaussian distribution is for features, but FIR impulse response or FFT result has more points, which is usually the sampling rate / frequency resolution. If the model is fitted for each frequency point, the number required is very large and the amount of calculation will be high. Appropriate feature dimensionality reduction can be considered. The following dimensionality reduction scheme is used:

[0077]

[0078] A total of 33 features are used to fit the Gaussian mixture model based on the data. During the fitting process, it is necessary to manually determine the actual vehicle label of the vehicle, or obtain the label by listening to the recording. Label is a term in machine learning. Data and the judgment corresponding to the data are called labels.

[0079] Model judgment

[0080] After the above steps, the subsequent process only needs to be done through feature extraction and then through the above fitting model to make a judgment to obtain the status of the vehicle off the line. The flowchart of the judgment process can be Figure 5 shown.

[0081] Model Update

[0082] Later, for the judged data or customer complaint data, the data can be re-listened and the labels can be re-confirmed. Then the training database is updated and finally the Gaussian mixture model is updated. The process diagram of model update and training can be as follows: Figure 6 shown.

[0083] This example proposes a method for off-line vehicle testing using the speaker microphone inside the vehicle. By fitting the time domain and frequency domain features and then fitting the Gaussian mixture model, the speaker microphone system of the whole vehicle off-line can be tested, overcoming the problem of difficult off-line vehicle testing:

[0084] (1) No additional microphones or speakers are required for separate testing;

[0085] (2) The microphone and speaker can be tested simultaneously;

[0086] (3) Save testing time.

[0087] Explanation for testing microphones and speakers at the same time: If all speakers have problems with the transfer function of a single microphone, the microphone is faulty. If a single pair of speakers has problems with the transfer function of microphones, the speaker is faulty.

[0088] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with this technology to understand the contents of the present invention and implement them accordingly. They cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for off-line detection of an automotive acoustic system, characterized in that: include: S1, Playback and Collection, using the speakers and microphones in the car to test the hybrid system composed of its own speakers and microphones. The system drives the speakers one by one to play the set signal, and synchronously controls the microphone to collect and obtain the recording data; S2, transfer function calculation, calculating the transfer function between the speaker and the microphone in S1, and extracting the features of the transfer function through time domain features and frequency domain features; S3. Model training: Model training is performed through unsupervised learning methods. A Gaussian model is fitted to normal samples as a discriminant model. The amount of computation is reduced by feature dimensionality reduction. Feature dimensionality reduction includes using the time difference to the first peak in the FIR time domain transfer function, the maximum value in the FIR time domain transfer function, and the FIR frequency domain transfer function as 1 / 3 octaves. During the model training and fitting process, the real car label is first determined manually or the label is obtained by listening to the recording. S4, judgment, judging by the trained discriminant model, extracting features, and then judging the status of the vehicle off the line by the discriminant model fitted in S3; It also includes a model update step, which involves re-listening to the judged data or customer complaint data, re-confirming the labels, and then updating the training database and finally updating the Gaussian mixture model.

2. The off-line detection method for an automotive acoustic system according to claim 1, characterized in that: The speaker in S1 generates a white noise signal.

3. The off-line detection method for an automotive acoustic system according to claim 1, characterized in that: When S1 is excited, a low-pass filter is performed to filter out high-frequency signals above 1 / 2 of the sampling rate set by the corresponding microphone.

4. The off-line detection method for an automotive acoustic system according to claim 1, characterized in that: In S2, the finite impulse response model estimation method is used to calculate the transfer function.

Images