Method for predicting acoustic performance of electronic parking brake system

Abstract

The invention relates to the technical field of braking systems, in particular to an acoustic performance prediction method for an electronic parking braking system. A method for predicting the acoustic performance of an electronic parking brake system comprises a first-stage transmission mechanism and is characterized by comprising the following specific steps: (1) collecting the sizes of parts influencing MGU noise in the first-stage transmission mechanism; (2) sample SPL calculation; (3), carrying out the loudness calculation of the Moore-Glasberg-Schlittenache, and carrying out the loudness calculation of the Moore-Glasberg-Schlittenache; (4) BP neural network structure algorithm; (5) genetic algorithm; and (6) designing a GA-BP fusion model. Compared with the prior art, the method for predicting the acoustic performance of the electronic parking brake system has the advantages that a plurality of key sizes, SPL (sound pressure level) and loudness value acoustic parameters of an MGU transmission system are collected, iterative learning is carried out by applying a genetic BP (GA-BP) neural network, and the MGU acoustic performance prediction method is obtained and used for later project development and optimization; the test time is reduced to a great extent, and the overall working efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of brake systems, and in particular to a method for predicting acoustic performance of an electronic parking brake system. Background Art

[0002] The Electronic Parking Brake (EPB) system is an electronic control system used to replace the traditional manual parking brake (handbrake), and is currently widely used in the passenger car field. Its common working scenarios are divided into two types: activating and releasing the parking brake. When the driver presses the EPB button, the activation and release of the parking brake instructions are sent to the system, and the motor drives the gear transmission system to execute the clamping and release instructions to realize the function. With the rapid development of the new energy market, end users and vehicle manufacturers have also put forward higher requirements for the acoustic performance of EPB during operation. As the main noise contributor when EPB is working, MGU (electromechanical brake actuator) is particularly important for quickly and accurately predicting its acoustic performance to evaluate the acoustic performance of the electronic parking brake system. In many years of experience in MGU noise optimization, it was found that several key dimensions of the transmission system may affect the acoustic performance. Summary of the Invention

[0003] To overcome the shortcomings of the existing technology, the present invention provides a method for predicting the acoustic performance of an electronic parking brake system. By collecting several key dimensions and SPL (sound pressure level) and loudness value acoustic parameters of the MGU powertrain, a genetic BP (GA-BP) neural network is used for iterative learning to obtain the MGU acoustic performance prediction method.

[0004] To achieve the above objectives, a method for predicting the acoustic performance of an electronic parking brake system is designed, which includes a first-stage transmission mechanism and is characterized by the following specific method:

[0005] (1) Collect the dimensions of the parts in the primary transmission mechanism that affect the MGU noise;

[0006] (2) Sample SPL calculation;

[0007] (3) Moore-Glasberg-Schlittenlache loudness calculation;

[0008] (4) BP neural network structure algorithm;

[0009] (5) Genetic algorithm;

[0010] (6) GA-BP fusion model design.

[0011] In the step (1), the dimensions of the components that affect the MGU noise include the distances A, B, and C between the motor output shaft and the center axis of the primary gear at different distances from the shaft top; the tooth width D of the primary transmission gear; and the gear accuracy grade E.

[0012] The A is the distance between the motor output shaft and the center axis of the first gear at 3 mm from the top of the shaft; B is the distance between the motor output shaft and the center axis of the first gear at 6 mm from the top of the shaft; C is the distance between the motor output shaft and the center axis of the first gear at 9 mm from the top of the shaft.

[0013] The process of step (2) is as follows:

[0014] S21, combining the part to be tested in step (1) with a standard caliper, and completing the release and clamping actions normally under the standard test voltage;

[0015] S22, using a microphone to collect sound signals and converting the sound signals into time domain digital signals through an analog-to-digital converter, the unit of which is pa;

[0016] S23, performs discrete STFT transform on the time domain digital signal:

[0017]

[0018] Where x(n) is the input signal; N is the frame length; N-1 is the frequency index, k = 0, 1, 2, ...; Δf is the frequency domain resolution, f = k * Δf, n = 0, 1, 2, ... N-1;

[0019] S24, performing A-weighted filtering according to discrete STFT transform and obtaining an A-weighted curve;

[0020] S25, calculate the A-weighted sound pressure level X A (f), the corresponding relationship between sound pressure P and sound pressure level SPL: SPL = 20lg (P / P0), where P is the actual sound pressure (Pa) and P0 = 2×10^-5Pa is the reference sound pressure. The sound pressure level is calculated based on X(f), and the corresponding frequency is attenuated according to A(f) to obtain X A (f), in dB, is the SPL value of the part under test.

[0021] The transfer function of the A-weighted filter is expressed in dB scale in the frequency domain as:

[0022]

[0023] The process of step (3) is as follows:

[0024] S31, after the sound signal passes through the outer ear and middle ear, the energy may increase or decrease accordingly. This is obtained by polynomial fitting based on the transfer function values ​​of the outer ear and middle ear transfer function at discrete frequencies in ISO 532-3;

[0025] S32, calculate the filter bandwidth ERB, ERB = 24.673 (0.004368f + 1), where f is the frequency in Hz; ERB is the bandwidth in Hz;

[0026] S33, calculate the filter center frequency ERB-number, ERB-number=21.3661g(0.004368f+1), and further calculate the excitation level;

[0027] S34, the excitation level is obtained, and after the sound signal is filtered by the basilar membrane, the characteristic loudness can be calculated:

[0028]

[0029] Where E is the energy level of the signal; E THRQ The energy level of the audible threshold is in dB; C is a fixed value of 0.046871; above 500Hz, E THRQ Take the fixed value 2.3067, G = 0.1, a = 0.2;

[0030] S35, sum to calculate the total loudness:

[0031] From 1.8ERB-number to 38.9ERB-number, with an interval of 0.1, a total of 372 filters are constructed; the shape of each filter is an exponential weight function W(g), where W(g) is the weighted value of the corresponding frequency component, and the calculation formula is: W(g) = (1 + pg)e (-pg) ;

[0032]

[0033] Where f1 is the signal frequency; f0 is the center frequency of the filter; L is the excitation level corresponding to the center frequency calculated in Section 2.5, and the calculation formula is: L = 10lg (E / E0), Where E is the signal energy; X(k) is the Fourier transform result of the signal, N is the length of the signal; E0=10e-12, the unit of the excitation level is dB, and p and g are dimensionless values ​​in the calculation process.

[0034] The process of step (4) is as follows:

[0035] S41, combining the part size data affecting MGU noise in step (1), and outputting the sound pressure level and loudness of the MGU acoustic test results;

[0036] S42, input data x i Scale to [0,1]: Among them, n i is the normalized data, i=1,2,……m; x is the original data; x min is the minimum value in the original data, x max is the maximum value in the original data;

[0037] S43, the normalized data is used to establish an MGU size-acoustic performance matrix, where the input of each layer is represented by I, the output is represented by O, the weight is represented by w, the threshold is represented by b, the superscript number represents the level, the subscript number letter represents the number of neurons in each layer, the input variable is X, m is the number of input layer variables, n is the number of hidden layer neurons, and l is the number of output variables;

[0038]

[0039] The hidden layer is: The output layer is:

[0040] When the network output With the expected output O e When they are not equal, there is an output error F, namely: O e is the expected output value;

[0041] Expand the error F to the hidden layer: Expand further to the input layer:

[0042]

[0043] Among them, the network error F is the weight of each layer and threshold function;

[0044] S44, back propagation is performed using weight and threshold adjustments proportional to the decrease in the error gradient; the hidden layer weight and threshold adjustment formula is: Input layer weight, threshold adjustment formula is: Among them, the negative sign indicates gradient descent, and the constant α is the learning rate;

[0045] S45, based on steps S41 to S44, the weights and threshold gradients are continuously updated, and the cycle is repeated until convergence is finally achieved, which means that the iterative learning of the cyclic neural grid is completed, and the MGU predicted SPL and loudness values ​​are finally obtained.

[0046] The process of step (5) is as follows:

[0047] S51, population initialization: Each individual is composed of four parts: the connection weight and threshold between the input layer and the hidden layer, and the connection weight and threshold between the hidden layer and the output layer: ω = I*H + H*O, θ = H + O, Len = ω + θ; where I is the number of input layer nodes, H is the number of hidden layer nodes, O is the number of output layer nodes, ω is the number of weights, θ is the number of thresholds, and Len is the individual encoding length;

[0048] S52, fitness function: The mean square error between the predicted output and the expected output is taken as the individual fitness value f, and the calculation formula is: Among them, n is the sum of the number of test sets and validation sets, y i is the actual output of the i-th sample, is the expected output of the i-th sample;

[0049] S53, selection method based on fitness ratio, the selection probability of each individual i is: Where N is the number of individuals in the population, f i is the fitness value of individual i, p i is the probability of individual i being selected;

[0050] S54, using real number crossover method, the kth chromosome α k The crossover operation with the lth chromosome α1 at position j is:

[0051] α kj =α kj (1-b)+α ij bα ij =α ij (1-b)+α kj b; where b is a random number between [0,1];

[0052] S55, select the jth gene of the i-th individual for mutation, the operation is as follows:

[0053] Among them, α max is α ij The maximum value, α min Gene α ij The lower bound of f(g) = r2(1-g / G max ) 2 , r2 is a random number, g is the current number of iterations, G max is the maximum number of evolutions, and r is a random number between [0, 1].

[0054] The process of step (6) is as follows:

[0055] S61, determining the BP network topology;

[0056] S62, initial weight, threshold length;

[0057] S63, performing the genetic algorithm of steps S51 to S55 to calculate the fitness value of each individual in the contemporary population, and the individual with the smallest fitness value is the optimal individual;

[0058] S64, perform BP neural network prediction in steps S41 to S45, and finally obtain a model and predicted data.

[0059] Compared with the existing technology, the present invention provides a method for predicting the acoustic performance of an electronic parking brake system. By collecting several key dimensions of the MGU powertrain and acoustic parameters such as SPL (sound pressure level) and loudness value, and using a genetic BP (GA-BP) neural network for iterative learning, the method is obtained. This method can be used for future project development and optimization, greatly reducing testing time and improving overall work efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 Schematic diagram of the center distance of the first-stage transmission mechanism.

[0061] Figure 2 This is a schematic diagram of the dimensions of the center distances A, B, and C between the first-stage gear shaft and the motor shaft.

[0062] Figure 3 Schematic diagram of the first-stage transmission gear tooth width and gear accuracy grade.

[0063] Figure 4 Flowchart for Moore loudness calculation.

[0064] Figure 5 This is a graph of the outer ear transfer function.

[0065] Figure 6 This is a graph of the middle ear transfer function.

[0066] Figure 7 For E THRQ Value diagram.

[0067] Figure 8 Schematic diagram of the relationship between a and G value.

[0068] Figure 9 This is the topological structure diagram of the BP neural network.

[0069] Figure 10 This is the flow chart of the GA-BP fusion model.

[0070] Figure 11Key dimensions vs. SPL, loudness meter for MGU.

[0071] Figure 12 Schematic diagram of SPL prediction results.

[0072] Figure 13 This is the loudness prediction result diagram. DETAILED DESCRIPTION

[0073] The present invention will be further described below with reference to the accompanying drawings.

[0074] like Figures 1 to 3 As shown, the sizes of parts that affect MGU noise are collected: the center distances A, B, and C of the first-stage transmission mechanism at different positions (i.e., the distances A, B, and C between the motor output shaft and the center axis of the first-stage gear at different distances from the shaft top. A is the distance between the motor output shaft and the center axis of the first-stage gear at 3 mm from the shaft top; B is the distance between the motor output shaft and the center axis of the first-stage gear at 6 mm from the shaft top; C is the distance between the motor output shaft and the center axis of the first-stage gear at 9 mm from the shaft top), the tooth width D of the first-stage transmission gear, and the gear accuracy grade E, a total of 5 key dimensions.

[0075] 1. Sample SPL calculation

[0076] (1) Combine the part to be tested with the standard caliper and make it release and clamp normally under the standard test voltage;

[0077] (2) Use a microphone to collect sound signals and convert them into time domain digital signals x(n) through an analog-to-digital converter (ADC), with the unit being pa;

[0078] (3) Perform discrete STFT transform on the time domain digital signal: x(n) is the input signal; N is the frame length (number of sampling points per frame); k = 0, 1, 2, ...; N-1 is the frequency index; f = k*Δf, Δf is the frequency domain resolution; X(k) unit is pa; n = 0, 1, 2, ... N-1;

[0079] (4) The A-weighted filter is a digital filter whose frequency response conforms to the A-weighted curve defined in the international standard IEC61672. The transfer function of the A-weighted filter in dB scale is expressed in the frequency domain as:

[0080]

[0081] (5) Calculate the A-weighted sound pressure level X A (f), the corresponding relationship between sound pressure P and sound pressure level SPL: SPL = 20lg (P / P0); where P is the actual sound pressure (Pa) P0 = 2×10^-5Pa is the reference sound pressure, the sound pressure level is calculated based on X(f), and the corresponding frequency is attenuated according to A(f) to obtain XA (f), in dB, is the SPL value of the part to be tested.

[0082] 3. Moore-Glasberg-Schlittenlache Loudness Calculation

[0083] Loudness is a psychoacoustic quantity that describes the intensity of sound perception by the human ear. Loudness calculation is based on the biological structure of the human ear, which processes sound signals. The human ear consists of three main parts: the outer ear, the middle ear, and the inner ear. Sound waves first travel to the outer ear, then through the auricle and the external auditory canal to the eardrum, causing it to vibrate and transmit the signal to the middle ear. As sound travels through the external auditory canal to the middle ear, the movement of the stapes causes changes in the fluid pressure within the cochlea, which in turn causes traveling waves to propagate along the basilar membrane. Nerve endings on this membrane transmit this information to the brain, resulting in hearing.

[0084] Moore's loudness calculation model is based on the principles of hearing and the attenuation characteristics of sound signals. It proposes a loudness calculation model that is closer to the real hearing characteristics of the human ear. The algorithm flow of this model is as follows: Figure 4 shown.

[0085] After the sound signal passes through the outer ear and middle ear, the energy may increase or decrease accordingly. Figure 5 and Figure 6 The curve is obtained by polynomial fitting based on the transfer function values ​​of the outer ear (free field) and middle ear transfer function at discrete frequencies in ISO532-3.

[0086] Specific gain reference Figure 5 and Figure 6 In the discrete STFT transform of the time-domain digital signal, the outer ear and middle ear transfer functions are increased or decreased based on X(f).

[0087] The excitation level represents the energy level of the sound signal in the cochlea after passing through the outer and middle ears. Moore loudness is calculated based on the ERB scale. Its mechanism is that the sound signal is transmitted to the basilar membrane, and a set of filters must be designed to accurately reflect the masking effect of the basilar membrane. The ERB scale was proposed based on the frequency response characteristics of the cochlear filter.

[0088] (1) Calculate the filter bandwidth ERB: ERB = 24.673(0.004368f + 1); where f is the frequency (Hz); ERB is the bandwidth (Hz);

[0089] (2) Calculate the filter center frequency ERB-number: ERB-number = 21.3661g (0.004368f + 1);

[0090] (3) Constructing filters To achieve more accurate calculations, a total of 372 filters are constructed, ranging from 1.8ERB-number to 38.9ERB-number, with an interval of 0.1. The shape of each filter is an exponential weight function W(g), where W(g) is the weight value of the corresponding frequency component, as shown below:

[0091] W(g)=(1+pg)e (-pg) ;

[0092]

[0093] Where f1 is the signal frequency; f0 is the center frequency of the filter; L is the excitation level corresponding to the center frequency calculated in Section 2.5, and the calculation formula is: L = 10lg (E / E0), Where E is the signal energy; X(k) is the Fourier transform result of the signal, N is the length of the signal; E0 = 10e-12, the unit of the excitation level is dB; p and g are dimensionless values ​​in the calculation process;

[0094] (4) After obtaining the excitation level according to the above figure, the sound signal can be filtered by the basilar membrane to calculate the characteristic loudness:

[0095]

[0096] Where E is the energy level of the signal; E THRQ The energy level of the audible threshold is in dB; C is a fixed value of 0.046871; above 500Hz, E THRQ Take the fixed value 2.3067, G equals 0.1, a equals 0.2; below 500Hz, E THRQ Take the fixed value Figure 7 As shown; G value is taken as E above 500Hz THRQ (2.3067) and E below 500Hz THRQ The value of a depends on the G value, and the relationship between the two is as follows: Figure 8 shown.

[0097] Sum and calculate the total loudness: Through the above calculation steps, we can get 372 characteristic loudness values. The total loudness is equal to the sum of these characteristic loudness values. Since the above calculations are for a single ear, to correctly obtain the loudness of the sound signal, we must multiply it by 2. That is:

[0098] 4. BP neural network structure

[0099] BP neural networks have strong nonlinear mapping capabilities and can perform adaptive learning. They use the backpropagation algorithm to correct errors and update weights, thereby achieving adaptive learning optimization. A BP neural network is similar to a nonlinear function, consisting of an input layer, hidden layers, and an output layer. The network's input and output are the independent and dependent variables of the function, respectively. The structure of a BP neural network directly affects its performance. Before network training, parameters such as the number of neurons in the input, hidden, and output layers, the transfer function, the learning function, and the target error value must be defined.

[0100] (1) Sample selection: The five key dimensions, the center distances A, B, and C of the primary transmission mechanism at different positions, the tooth width D of the primary transmission gear, and the gear accuracy grade E, are used as input data for the prediction model, and the sound pressure level and loudness of the MGU acoustic test results are used as output variables;

[0101] (2) Sample normalization: transform the input data x i Scaling to [0,1], we get Among them, n i is the normalized data, i=1,2,……m, x is the original data; x min is the minimum value in the original data, x max is the maximum value in the original data. The normalized data is used to establish the MGU size-acoustic performance matrix;

[0102] (3) Forward propagation process: The input of each layer is represented by I, the output is represented by O, the weight is represented by w, the threshold is represented by b, the superscript number represents the level, the subscript number letter represents the number of neurons in each layer, and f() in the formula is the activation function. The input variable is X, m is the number of input layer variables, n is the number of hidden layer neurons, and l is the number of output variables. The BP neural network topology structure is as follows: Figure 9 As shown;

[0103]

[0104] The hidden layer is:

[0105] Output layer, output node gets the output value Right now:

[0106]

[0107] When the network output With the expected output O e When they are not equal, there is an output error F, namely:

[0108] Among them, O e is the expected output value;

[0109] Expanding the error F to the hidden layer gives:

[0110]

[0111] Further expansion to the input layer:

[0112]

[0113] Among them, the network error F is the weight of each layer

[0114] and threshold function, thus adjusting the weights and thresholds can change the error;

[0115] (4) Back propagation process: The back propagation process is to continuously reduce the error, so the adjustment amount of the weight and threshold should be proportional to the decrease of the error gradient. The hidden layer weight and threshold adjustment formula is: Input layer weight, threshold adjustment formula: Here, the negative sign indicates gradient descent, and the constant α is the learning rate.

[0116] Based on the above theory, weights and threshold gradients are continuously updated in a cycle until convergence is achieved (when the MSE falls below a preset threshold, the number of iterations reaches a preset value, the gradient norm of the weight update falls below a certain threshold, and the weight update amount falls below the preset threshold). This indicates that iterative learning of the neural network is complete, ultimately obtaining the MGU predicted SPL and loudness values. However, using the BP neural network algorithm alone is prone to falling into local optimality. To improve this problem, a genetic algorithm is used before the calculation to find the optimal weights and thresholds.

[0117] 5. Genetic Algorithm

[0118] The greatest advantage of genetic algorithms lies in their global optimization capabilities. The algorithm is less likely to fall into local extremes, and even under unfavorable conditions such as discontinuities in the fitness function, it is still more likely to find the optimal solution than other algorithms. Furthermore, genetic algorithms are highly scalable and can be easily combined with other algorithmic models. This article combines genetic algorithms with neural networks, and the final results clearly demonstrate the superior performance of this combination.

[0119] (1) Population initialization: There are two main ways of encoding: binary encoding and real number encoding. In this paper, in order to improve the computational complexity of the GA algorithm and improve efficiency, real number encoding is used for individual encoding. Each individual is composed of four parts: the connection weights and thresholds between the input layer and the hidden layer, and the connection weights and thresholds between the hidden layer and the output layer: ω = I*H + H*O; θ = H + O; Len = ω + θ; where I is the number of input layer nodes, H is the number of hidden layer nodes, O is the number of output layer nodes, ω is the number of weights, θ is the number of thresholds, and Len is the individual encoding length;

[0120] (2) Fitness function: During training, we hope that the error between the predicted value and the expected value is as small as possible, so the mean square error between the predicted output and the expected output is used as the individual fitness value f. The calculation formula (selecting the mean square error loss function MSE) is: Among them, n is the sum of the number of test sets and validation sets, y i is the actual output of the i-th sample, is the expected output of the i-th sample;

[0121] (3) Selection: The selection operation of the present invention is based on the fitness ratio selection method, and the selection probability of each individual i is: Where N is the number of individuals in the population, f i is the fitness value of individual i, p i

[0122] is the probability of individual i being selected;

[0123] (4) Crossover: The present invention adopts real number crossover method, and the kth chromosome α k and chromosome α l The crossover operation at position j is:

[0124] α kj =α kj (1-b)+α ij b

[0125] α ij =α ij (1-b)+α kj b

[0126] Where b is a random number between [0,1];

[0127] (5) Mutation: Select the jth gene of the i-th individual for mutation. The operation is as follows:

[0128] Among them, α max is α ij The maximum value, α min Gene αij The lower bound of f(g) = r2(1-g / G max ) 2 , r2 is a random number, g is the current number of iterations, G max is the maximum number of evolutions, and r is a random number between [0, 1].

[0129] 6. GA-BP Fusion Model Design

[0130] like Figure 10 As shown in the figure, the GA-BP fusion model process is as follows:

[0131] (1) Determine the BP network topology;

[0132] (2) Initial weight, threshold length;

[0133] (3) Calculate the fitness value of each individual in the contemporary population by performing the above genetic algorithm, and the individual with the smallest fitness value is the optimal individual;

[0134] (4) Carry out the BP neural network prediction mentioned above, finally obtain the model and obtain the predicted data.

[0135] Example:

[0136] like Figure 11 As shown in the table, 70% of the data is used as training set and the rest is test set.

[0137] GA parameter settings: initial population 30, maximum evolutionary generations 50, crossover probability 0.8, mutation probability 0.2.

[0138] BP neural network parameters:

[0139] (1) Network structure: The input layer has 5 neurons, the output layer has 2 neurons, and the hidden layer nodes are calculated according to the empirical formula, where a is a constant with a value range of 1 to 10. In this embodiment, 10 neurons are selected for the hidden layer.

[0140] (2) Activation function: The hidden layer activation function uses the tansig function, and the output layer activation function uses the linear activation function (purelin);

[0141] (3) Training function: The training function is trainlm, which uses the Levenberg-Marquardt algorithm;

[0142] (4) Loss function: Select the mean square error loss function MSE, Among them, y i is the true value of the i-th sample, is the predicted value of the i-th sample, n is the number of samples;

[0143] (5) Learning rate, maximum number of training times: learning rate α = 0.01, maximum number of training times 1000 times;

[0144] (6) Prediction results: The prediction results of the BP and GA-BP algorithm models were compared, showing that the GA-BP algorithm has a higher degree of fit. Figure 12 For SPL prediction results, Figure 13 The loudness prediction result.

Claims

1. A method for predicting the acoustic performance of an electronic parking brake system, comprising a first-stage transmission mechanism, characterized in that: The specific method is as follows: (1) Collect the dimensions of the parts in the primary transmission mechanism that affect the MGU noise; (2) Sample SPL calculation; (3) Moore-Glasberg-Schlittenlache loudness calculation; (4) BP neural network structure algorithm; (5) Genetic algorithm; (6) GA-BP fusion model design.

2. The method for predicting acoustic performance of an electronic parking brake system according to claim 1, characterized in that: In the step (1), the dimensions of the components that affect the MGU noise include the distances A, B, and C between the motor output shaft and the center axis of the primary gear at different distances from the shaft top; the tooth width D of the primary transmission gear; and the gear accuracy grade E.

3. The method for predicting acoustic performance of an electronic parking brake system according to claim 2, wherein: The A is the distance between the motor output shaft and the center axis of the first gear at 3 mm from the top of the shaft; B is the distance between the motor output shaft and the center axis of the first gear at 6 mm from the top of the shaft; C is the distance between the motor output shaft and the center axis of the first gear at 9 mm from the top of the shaft.

4. The method for predicting acoustic performance of an electronic parking brake system according to claim 1, wherein: The process of step (2) is as follows: S21, combining the part to be tested in step (1) with a standard caliper, and completing the release and clamping actions normally under the standard test voltage; S22, using a microphone to collect sound signals and converting the sound signals into time domain digital signals through an analog-to-digital converter, the unit of which is pa; S23, performs discrete STFT transform on the time domain digital signal: Where x(n) is the input signal; N is the frame length; N-1 is the frequency index, k = 0, 1, 2, ...; Δf is the frequency domain resolution, f = k * Δf, n = 0, 1, 2, ... N-1; S24, performing A-weighted filtering according to discrete STFT transform and obtaining an A-weighted curve; S25, calculate the A-weighted sound pressure level X A (f), the corresponding relationship between sound pressure P and sound pressure level SPL: SPL = 20lg (P / P0), where P is the actual sound pressure (Pa) and P0 = 2×10^-5Pa is the reference sound pressure. The sound pressure level is calculated based on X(f), and the corresponding frequency is attenuated according to A(f) to obtain X A (f), in dB, is the SPL value of the part under test.

5. The method for predicting acoustic performance of an electronic parking brake system according to claim 3, characterized in that: The transfer function of the A-weighted filter is expressed in dB scale in the frequency domain as:

6. The method for predicting acoustic performance of an electronic parking brake system according to claim 1, characterized in that: The process of step (3) is as follows: S31, after the sound signal passes through the outer ear and middle ear, the energy may increase or decrease accordingly. This is obtained by polynomial fitting based on the transfer function values ​​of the outer ear and middle ear transfer function at discrete frequencies in ISO 532-3; S32, calculate the filter bandwidth ERB, ERB = 24.673 (0.004368f + 1), where f is the frequency in Hz; ERB is the bandwidth in Hz; S33, calculate the filter center frequency ERB-number, ERB-number=21.3661g(0.004368f+1), and further calculate the excitation level; S34, the excitation level is obtained, and after the sound signal is filtered by the basilar membrane, the characteristic loudness can be calculated: Where E is the energy level of the signal; E THRQ The energy level of the audible threshold is in dB; C is a fixed value of 0.046871; above 500Hz, E THRQ Take the fixed value 2.3067, G = 0.1, a = 0.2; S35, sum to calculate the total loudness:

7. The method for predicting acoustic performance of an electronic parking brake system according to claim 6, characterized in that: From 1.8ERB-number to 38.9ERB-number, with an interval of 0.1, a total of 372 filters are constructed; the shape of each filter is an exponential weight function W(g), where W(g) is the weight value of the corresponding frequency component, and the calculation formula is: W(g)=(1+pg)e (-pg) ; Where f1 is the signal frequency; f0 is the center frequency of the filter; L is the excitation level corresponding to the center frequency calculated in Section 2.5, and the calculation formula is: L = 10lg (E / E0), Where E is the signal energy; X(k) is the Fourier transform result of the signal, N is the length of the signal; E0=10e-12, the unit of the excitation level is dB, and p and g are dimensionless values ​​in the calculation process.

8. The method for predicting acoustic performance of an electronic parking brake system according to claim 1, characterized in that: The process of step (4) is as follows: S41, combining the part size data affecting MGU noise in step (1), and outputting the sound pressure level and loudness of the MGU acoustic test results; S42, input data x i Scale to [0,1]: Among them, n i is the normalized data, i=1,2,……m; x is the original data; x min is the minimum value in the original data, x max is the maximum value in the original data; S43, the normalized data is used to establish an MGU size-acoustic performance matrix, where the input of each layer is represented by I, the output is represented by O, the weight is represented by w, the threshold is represented by b, the superscript number represents the level, the subscript number letter represents the number of neurons in each layer, the input variable is X, m is the number of input layer variables, n is the number of hidden layer neurons, and l is the number of output variables; The hidden layer is: The output layer is: When the network output With the expected output O e When they are not equal, there is an output error F, namely: O e is the expected output value; Expand the error F to the hidden layer: Expand further to the input layer: Among them, the network error F is the weight of each layer and threshold function; S44, back propagation is performed using weight and threshold adjustments proportional to the decrease in the error gradient; the hidden layer weight and threshold adjustment formula is: Input layer weight, threshold adjustment formula is: Among them, the negative sign indicates gradient descent, and the constant α is the learning rate; S45, based on steps S41 to S44, the weights and threshold gradients are continuously updated, and the cycle is repeated until convergence is finally achieved, which means that the iterative learning of the cyclic neural grid is completed, and the MGU predicted SPL and loudness values ​​are finally obtained.

9. The method for predicting acoustic performance of an electronic parking brake system according to claim 1, characterized in that: The process of step (5) is as follows: S51, population initialization: Each individual is composed of four parts: the connection weight and threshold between the input layer and the hidden layer, and the connection weight and threshold between the hidden layer and the output layer: ω = I*H + H*O, θ = H + O, Len = ω + θ; where I is the number of input layer nodes, H is the number of hidden layer nodes, O is the number of output layer nodes, ω is the number of weights, θ is the number of thresholds, and Len is the individual encoding length; S52, fitness function: The mean square error between the predicted output and the expected output is taken as the individual fitness value f, and the calculation formula is: Among them, n is the sum of the number of test sets and validation sets, y i is the actual output of the i-th sample, is the expected output of the i-th sample; S53, selection method based on fitness ratio, the selection probability of each individual i is: Where N is the number of individuals in the population, f i is the fitness value of individual i, p i is the probability of individual i being selected; S54, using real number crossover method, the kth chromosome α k The crossover operation with the lth chromosome α1 at position j is: a kj =a kj (1-b)+a ij b α ij =α ij (1-b)+α kj b; where b is a random number between [0,1]; S55, select the jth gene of the i-th individual for mutation, the operation is as follows: Among them, α max is α ij The maximum value, α min Gene α ij The lower bound of f(g) = r2(1-g / G max ) 2 , r2 is a random number, g is the current number of iterations, G max is the maximum number of evolutions, and r is a random number between [0, 1].

10. The method for predicting acoustic performance of an electronic parking brake system according to claim 1, characterized in that: The process of step (6) is as follows: S61, determining the BP network topology; S62, initial weight, threshold length; S63, performing the genetic algorithm of steps S51 to S55 to calculate the fitness value of each individual in the contemporary population, and the individual with the smallest fitness value is the optimal individual; S64, perform BP neural network prediction in steps S41 to S45, and finally obtain a model and predicted data.

Images