Sunroof rail torsion resistance performance evaluation method based on current ripple and vibration characteristics

Abstract

The application discloses a sunroof guide rail torsional performance evaluation method based on current ripple and vibration characteristics, and belongs to the technical field of automobile part detection; a driving motor is connected with the sunroof guide rail through a shaft coupling to output periodic torsional motion; a motor current ripple signal and a bearing seat vibration signal are synchronously collected and digitized through a multi-channel data acquisition card; the current ripple is sequentially subjected to first-stage high-pass filtering, second-stage adaptive notch filtering and third-stage five-point cubic smoothing, 12-dimensional characteristic parameters and vibration compensation amounts are extracted; 8-dimensional core characteristic parameters are screened through a mutual information method; a genetic algorithm-optimized back propagation neural network is constructed as a characteristic stress mapping function; the 8-dimensional core characteristic parameters of a to-be-tested guide rail are input into a current ripple torsional performance mapping model, torsional stiffness, torsional strength and fatigue life evaluation values are output, and non-destructive, rapid and accurate evaluation is realized.

Description

Technical Field

[0001] This invention belongs to the technical field of automotive parts testing, specifically relating to a method for evaluating the torsional performance of sunroof guide rails based on current ripple and vibration characteristics. Background Technology

[0002] With the development of the automotive industry, the reliability and durability of sunroof systems are becoming increasingly important. As a core load-bearing component, the sunroof rails must withstand cyclic torsional loads caused by their own weight, road vibrations, temperature changes, and opening / closing operations. Their torsional resistance directly affects the sunroof's service life and safety. Therefore, it is necessary to evaluate their torsional stiffness, torsional strength, and fatigue life.

[0003] In the field of automotive sunroof testing technology, Chinese invention patent application CN112629877A discloses an automotive sunroof testing method and system. It obtains a fatigue load rainflow matrix and a sunroof fatigue damage matrix based on typical road driving data, establishes the correspondence between sunroof damage and fatigue load, determines the target load accordingly, and inputs the equivalent load to the vehicle test bench to complete the vehicle-level acceleration durability test, so as to identify sunroof failure problems in advance.

[0004] However, the aforementioned existing technologies still have the following shortcomings: First, while this solution can assess fatigue damage to sunroofs, it relies on vehicle-level bench testing and the collection of typical road driving data, resulting in a long testing cycle. Furthermore, it requires the sunroof to be assembled in a complete vehicle or on a vehicle bench, making it difficult to perform rapid online testing of individual sunroof rails at the end of the parts production line. Additionally, this solution does not involve current ripple feature extraction, and therefore cannot utilize the load torque modulation information inherent in the drive motor's operation to evaluate the torsional performance of the sunroof rail.

[0005] Furthermore, existing current ripple analysis is typically used for motor condition monitoring and fault diagnosis, treating current ripple as an indicator of the motor's own condition. Under controlled torsional testing conditions, the drive motor's own state is essentially constant, and the changes in current ripple characteristics mainly originate from the modulation of the motor's load torque through the transmission chain by the mechanical response of the tested sunroof guide rail under torsional load. Current technology has not yet established an effective correlation between this current ripple variation and the torsional resistance performance of the tested sunroof guide rail, nor does it possess a quantitative evaluation method that integrates it with vibration characteristics.

[0006] Therefore, how to quickly and accurately assess the torsional performance of the sunroof guide rail without damaging it, in order to meet the batch testing requirements of the production line, has become a technical problem that needs to be solved in this field. Summary of the Invention

[0007] The purpose of this invention is to provide a method for evaluating the torsional performance of sunroof guide rails based on current ripple and vibration characteristics, which can effectively solve the problems in the background art.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for evaluating the torsional performance of sunroof guide rails based on current ripple and vibration characteristics includes: connecting a drive motor to the sunroof guide rail via a coupling on a dedicated torsion test bench for sunroof guide rails, with the drive motor outputting periodic torsional motion; synchronously acquiring the current ripple signal of the drive motor and the vibration signal of the drive motor bearing housing, and digitizing the current ripple signal of the drive motor and the vibration signal of the drive motor bearing housing through a multi-channel data acquisition card.

[0009] The digitized drive motor current ripple signal is subjected to a first-level high-pass filter, a second-level adaptive notch filter, and a third-level five-point triple smoothing process to extract 12-dimensional feature parameters and extract vibration compensation from the drive motor bearing housing vibration signal.

[0010] The mutual information method is used to calculate the mutual information value between each feature parameter in the 12-dimensional feature parameters and the torsional stress and strain. The core feature parameters are selected according to the mutual information threshold. A feature stress mapping function is constructed with an 8-dimensional core feature vector as input and the estimated torsional stress and strain as output. The feature stress mapping function is implemented by a backpropagation neural network optimized by a genetic algorithm.

[0011] The eight core feature parameters of the sunroof guide rail under test are input into the trained current ripple torsional performance mapping model, and the torsional stiffness evaluation value, torsional strength evaluation value, and fatigue life evaluation value are output.

[0012] Furthermore, the drive motor current ripple signal is acquired by a Hall effect closed-loop current sensor, which is installed on any one phase of the three-phase power supply line of the drive motor using a through-hole mounting method.

[0013] The vibration signal of the drive motor bearing housing is collected by a piezoelectric accelerometer, which is fixed to the end face of the drive motor bearing housing by a threaded connection.

[0014] The acquisition trigger terminals of the Hall effect closed-loop current sensor and the piezoelectric accelerometer are connected to the same trigger source, so that the current ripple signal of the drive motor and the vibration signal of the drive motor bearing housing are kept in sync in terms of timing, and the timestamps are aligned with each other.

[0015] Furthermore, the digitized drive motor current ripple signal is subjected to a first-stage high-pass filter, a second-stage adaptive notch filter, and a third-stage five-point triple smoothing process in sequence, including: using a second-order Butterworth high-pass filter for the first-stage high-pass filter, with the cutoff frequency set to 400Hz.

[0016] A lattice-type adaptive filter structure is used for the second stage of adaptive notch filtering. The center frequency is set to 50Hz and the bandwidth is set to 200Hz. The filter coefficients are updated according to the minimum mean square criterion of the error signal so that the notch center frequency follows the change of the power grid frequency.

[0017] Based on the principle of cubic spline interpolation, cubic polynomial fitting is performed on 5 consecutive sampling points, followed by a third-level five-point cubic smoothing.

[0018] Furthermore, the vibration compensation amount is extracted from the vibration signal of the drive motor bearing housing, including: calculating the ratio of the energy of the power spectrum of the drive motor bearing housing vibration signal at the ripple characteristic frequency to the energy of the entire frequency band, wherein the ripple characteristic frequency includes 1kHz, 2kHz and 3kHz.

[0019] The vibration compensation amount is extracted after undergoing the same first-stage high-pass filtering, second-stage adaptive notch filtering, and third-stage five-point triple smoothing process as the drive motor current ripple signal.

[0020] Mechanical vibration of the shaft system changes the air gap length of the motor, causing changes in air gap magnetic reluctance, which in turn changes the air gap magnetic flux density, and consequently, the current ripple signal of the drive motor changes accordingly.

[0021] Furthermore, the mutual information method is used to calculate the mutual information value between each feature parameter in the 12-dimensional feature parameters and the torsional stress and strain, and the core feature parameters are selected according to the mutual information threshold. This includes: calculating the mutual information value between each feature parameter in the 12-dimensional feature parameters and the torsional stress and strain, and the formula for calculating the mutual information value is as follows:

[0022] in, For characteristic parameters With torsional stress or strain The mutual information value between them, in bits. For the feature parameters, any dimension from the 12-dimensional feature parameters can be selected. Torsional stress or strain; For feature parameters With torsional stress or strain The joint probability density function, For feature parameters The marginal probability density function, Torsional stress or strain The marginal probability density function.

[0023] The mutual information threshold is set to 0.85, and feature parameters with mutual information values ​​greater than 0.85 are selected into the core feature parameter set.

[0024] Eight core feature parameters were selected from the 12 feature parameters. These eight core feature parameters are: ripple fundamental amplitude, ripple second harmonic amplitude, total harmonic distortion rate, ripple amplitude standard deviation, peak value of harmonic power spectral density, slope of amplitude envelope, frequency stability index, and vibration compensation.

[0025] Furthermore, in the backpropagation neural network optimized by the genetic algorithm: the population size of the genetic algorithm is set to 50, the crossover probability is set to 0.7, the mutation probability is set to 0.01, the maximum number of iterations is set to 200 generations, the individual encoding method is real number encoding, and the fitness function is the root mean square error of the backpropagation neural network on the validation set.

[0026] The backpropagation neural network adopts a three-layer fully connected feedforward neural network topology, with 8 nodes in the input layer, 15 nodes in the hidden layer, and the hyperbolic tangent activation function for the hidden layer neurons. The output layer has 2 nodes and uses a linear activation function.

[0027] The goodness of fit of the characteristic stress mapping function on the validation set is greater than or equal to 0.92.

[0028] Furthermore, the current ripple anti-torsion performance mapping model adopts a three-layer backpropagation neural network with 8 nodes in the input layer, 15 nodes in the hidden layer, hyperbolic tangent activation function for the hidden layer neurons, 3 nodes in the output layer, and linear activation function for the output layer.

[0029] The weight matrix W1 from the input layer to the hidden layer has a dimension of 15×8, and the weight matrix W2 from the hidden layer to the output layer has a dimension of 3×15.

[0030] Each of the 1-dimensional feature parameters in the 8-dimensional core feature parameters is independently standardized by z-score, and there is no cross-coupling between the dimensions.

[0031] Furthermore, during testing on a dedicated torsion test bench for sunroof guide rails, the drive motor speed and torsional load remained stable, and the motor's own state remained essentially constant.

[0032] The change in current ripple characteristics in the drive motor current ripple signal originates from the mechanical response of the tested sunroof guide rail under torsional load, which modulates the motor load torque through the transmission chain. The mechanical response includes local stiffness changes caused by the evolution of microscopic defects in the material.

[0033] The torsional stress and strain estimates were calibrated and verified using stress and strain data obtained from traditional torsional fatigue tests.

[0034] Furthermore, the system outputs torsional stiffness assessment values, torsional strength assessment values, and fatigue life assessment values, including: the torsional stiffness assessment value is output as a percentage of the average torsional stiffness of the same batch of qualified sunroof guide rails, with an output range of 70% to 130%.

[0035] The torsional strength assessment value is output in four grades: excellent, good, medium, and poor. Excellent corresponds to a torsional strength percentage greater than or equal to 95%, good corresponds to a torsional strength percentage greater than or equal to 90% and less than 95%, medium corresponds to a torsional strength percentage greater than or equal to 85% and less than 90%, and poor corresponds to a torsional strength percentage less than 85%.

[0036] The fatigue life assessment value is output in hours, with a prediction range of 0 to 10,000 hours.

[0037] A testing database containing no less than 5,000 historical testing records was established. An autoregressive integral moving average model was used to predict the residual service life of the guide rail. The input of the autoregressive integral moving average model was the torsional performance evaluation results of the five most recent consecutive tests, and the output was the predicted value of the residual service life. The prediction lead was 1,000 hours to 5,000 hours.

[0038] Furthermore, the training set of the current ripple torsional performance mapping model shall have no less than 200 sets of samples and no less than 50 sets of samples. The neural network shall be trained using leave-one cross-validation.

[0039] After the current ripple torsional resistance performance mapping model is trained, its generalization performance is verified using an independent test set with no less than 50 test set samples. The generalization error of the current ripple torsional resistance performance mapping model is less than or equal to 5%.

[0040] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention enables non-destructive rapid online testing. It does not require assembling the sunroof guide rail onto the vehicle frame, nor does it rely on destructive torsional fatigue testing. Instead, it obtains the drive motor current ripple signal and bearing housing vibration signal through controlled periodic torsional motion on a dedicated torsional testing bench, thereby completing the torsional performance evaluation. The single-piece testing time is less than or equal to 45 seconds, which can meet the batch testing needs at the end of the production line.

[0041] 2. This invention can improve the accuracy of assessment and the ability to resist interference. It performs three-level preprocessing on the current ripple signal, which improves the signal-to-noise ratio to more than 35 dB. At the same time, it introduces vibration compensation to suppress mechanical vibration coupling interference, and uses the mutual information method to screen 8-dimensional core feature parameters. Then, it establishes a nonlinear mapping between current ripple characteristics and torsional stress and strain through a backpropagation neural network optimized by a genetic algorithm. This makes the torsional stiffness assessment error less than or equal to 4.2%, the accuracy rate of torsional strength grade determination reach 96.5%, the fatigue life prediction error less than or equal to 5.1%, and the misjudgment rate caused by vibration interference reduced by about 30%.

[0042] 3. This invention realizes the formation of a multi-dimensional assessment and life management closed loop, which can simultaneously output torsional stiffness assessment value, torsional strength assessment value and fatigue life assessment value, and store the assessment results in a test database of no less than 5,000 historical test records. Furthermore, it uses an autoregressive integral moving average model to predict the residual service life, with a prediction lead of 1,000 to 5,000 hours, thereby expanding a single torsional performance test into full life cycle quality tracking and early warning. Attached Figure Description

[0043] Figure 1 A schematic diagram of the overall technical solution for the evaluation method of torsional performance of sunroof guide rails based on current ripple and vibration characteristics; Figure 2 A schematic diagram of the core principle framework for the fusion analysis of current ripple and vibration characteristics; Figure 3 The logical flowchart for current ripple signal acquisition and vibration feature extraction is shown below. Figure 4 This is a schematic diagram of the interaction and data flow between the multi-source sensor signal fusion and torsional performance evaluation and judgment module. Figure 5 This diagram illustrates the technical performance comparison between this method and existing single-feature evaluation methods in terms of evaluation accuracy. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the appendix. Figure 1 To be continued Figure 5 The present invention will be further described in detail below with reference to specific embodiments.

[0045] The invention provides a method for evaluating the torsional performance of sunroof guide rails based on current ripple and vibration characteristics. The method collects the current signal of the drive motor in real time, extracts the multidimensional characteristic parameters of the current ripple, and performs coupling compensation by combining the vibration signal collected synchronously by the vibration sensor. Based on the nonlinear mapping algorithm, a quantitative mapping relationship between the current ripple characteristics and the torsional performance index is constructed, and the torsional stiffness, torsional strength and fatigue life of the sunroof guide rail are evaluated in a non-contact manner.

[0046] The overall technical architecture comprises a signal acquisition layer, a signal processing layer, a feature extraction layer, a model inference layer, and a data management layer. The signal acquisition layer performs synchronous acquisition and analog-to-digital conversion of current and vibration signals. The signal processing layer performs filtering, noise reduction, and power frequency interference suppression. The feature extraction layer extracts multi-dimensional ripple characteristic parameters and vibration compensation features. The model inference layer inputs the feature parameters into the mapping model and outputs the evaluation results. The data management layer stores historical detection data and performs residual lifetime prediction. Each layer transmits information through a standardized data interface, forming an evaluation closed loop. Based on the above overall technical architecture, the embodiments of this application are described in detail, and the specific implementation process is as follows.

[0047] It should be noted that this application differs fundamentally from existing technologies that utilize current ripple for motor fault diagnosis. In existing technologies, current ripple is typically considered an indicator signal of the motor's own condition, such as brush wear, bearing damage, or winding faults.

[0048] In this application, under controlled torsion test conditions, the drive motor speed and torsional load remain stable, the motor's own state is basically constant, and the change in current ripple characteristics mainly comes from the mechanical response of the tested sunroof guide rail under torsional load, which modulates the motor load torque through the transmission chain.

[0049] This mechanical response includes local stiffness changes caused by the evolution of microscopic defects in the material. Therefore, this application uses the current ripple signal as a response signal of the mechanical performance of the tested component, rather than an indication signal of the motor's own state. Under controlled operating conditions, there is a statistical correlation between the current ripple characteristics and the torsional stress-strain of the sunroof guide rail. This correlation is calibrated and verified using stress-strain data obtained from traditional torsional fatigue tests. Based on this statistical correlation, the torsional performance of the sunroof guide rail can be indirectly evaluated through the current ripple characteristics.

[0050] First, step S1 in the embodiment of this application is described: the current ripple signal of the drive motor and the vibration signal of the drive motor bearing housing are simultaneously acquired on the special torsion test bench for the sunroof guide rail, and used as input for subsequent current ripple feature extraction, vibration coupling compensation and anti-torsion performance mapping model inference, as detailed below.

[0051] In step S101, the sunroof guide rail is installed on a dedicated torsion test bench. The drive motor is connected to the sunroof guide rail via a coupling, driving the sunroof guide rail to generate periodic torsional motion. The input consists of the sunroof guide rail, the dedicated torsion test bench, and the drive motor. The processing method is to establish a mechanical connection between the drive motor and the sunroof guide rail via a coupling, and the drive motor outputs periodic torsional motion. The output is that the sunroof guide rail is in a controlled periodic torsional working condition.

[0052] Step S102: The Hall effect type closed-loop current sensor is installed in the power supply circuit of the drive motor. The installation position is selected on any one phase of the three-phase power supply line of the drive motor, using a through-hole installation method. The bandwidth of the Hall effect type closed-loop current sensor is 0Hz to 100kHz, the current measurement range is 0A to 50A, and the conversion accuracy is 0.1%.

[0053] The input is the phase current of the drive motor, with the operating current ranging from 0A to 50A, and the highest ripple frequency not exceeding 10kHz. The processing method is non-contact measurement using a Hall effect closed-loop current sensor. The Hall effect closed-loop current sensor performs non-contact current measurement based on the Hall effect, and the closed-loop design ensures measurement linearity and temperature stability. The output is a voltage signal proportional to the measured phase current, and the deviation of the measured current value from the true value does not exceed 0.1% within the range of 0A to 50A and 0Hz to 100kHz.

[0054] Step S103: The sampling frequency of the sample-and-hold circuit is set to 20kHz, and the quantization bit depth is set to 16bit. The input is the highest frequency of the current ripple signal to be acquired. ,Pick The processing method involves determining the sampling frequency based on the Nyquist sampling theorem. The formula is Substitute get Therefore, set Quantization bits Binary quantization series Substituting into the equation yields... The output has a sampling frequency of 20kHz to meet the current ripple acquisition requirements from 0Hz to 10kHz, and 16-bit quantization corresponds to 65536 levels of quantization.

[0055] in, The sampling frequency of the sample-and-hold circuit is expressed in Hz. The highest frequency of the current ripple signal to be acquired is expressed in Hz. The quantization bit depth is expressed in bits. It is a quantized series, dimensionless.

[0056] Step S104: The vibration sensor is a piezoelectric accelerometer with a sensitivity of 100mV / g and a frequency response range of 0.5Hz to 10kHz. It is installed on the end face of the drive motor bearing housing and fixed via a threaded connection. The 100mV / g sensitivity and the 0.5Hz to 10kHz frequency range match the vibration characteristic frequency band of the sunroof guide rail drive motor, covering the vibration frequency components generated during the torsion of the sunroof guide rail.

[0057] The vibration sensor and the Hall effect closed-loop current sensor are synchronously triggered using the same trigger source. The inputs are the current ripple voltage signal output from the Hall effect closed-loop current sensor and the vibration voltage signal output from the piezoelectric accelerometer.

[0058] The processing method involves using a piezoelectric accelerometer to acquire the vibration acceleration signal from the end face of the drive motor bearing housing. The sensor is fixed to the end face of the drive motor bearing housing via a threaded connection. The acquisition trigger terminals of the two sensors are connected to the same trigger source, so that the two signals start sampling at the same time and share the same time reference. The output is a voltage signal proportional to the vibration acceleration, covering the frequency band from 0.5Hz to 10kHz. The current ripple signal is consistent with the vibration signal in timing, and the timestamps are aligned with each other.

[0059] Step S105: The synchronously acquired current and vibration signals are digitized through the analog input port of the multi-channel data acquisition card. The data acquisition card has an embedded anti-aliasing filter with a cutoff frequency of... Set sampling frequency 1 / 2 of.

[0060] The input consists of synchronously acquired analog current ripple and vibration signals, as well as the sampling frequency. The processing method involves performing analog-to-digital conversion on the two analog signals through the analog input port of the multi-channel data acquisition card, enabling the built-in anti-aliasing filter of the data acquisition card, and following... Set the cutoff frequency and substitute it. get .

[0061] The output consists of frequency components higher than 10kHz, which are attenuated by an anti-aliasing filter. The output is a digitized current signal and a vibration signal. The digitized current signal and vibration signal are arranged into data frames according to the same timestamp. Each data frame contains the current value and vibration acceleration value at the same sampling time.

[0062] in, This is the cutoff frequency of the anti-aliasing filter, in Hz. This is the sampling frequency of the sample-and-hold circuit, in Hz.

[0063] At this point, step S1 outputs the synchronously acquired current and vibration signals, which serve as input data for subsequent current ripple characteristic parameter extraction, vibration compensation, and torsional performance mapping model construction.

[0064] After the above steps are completed, the following describes step S2 in this embodiment of the application: extraction of current ripple feature parameters. The synchronous current signal and vibration signal output in step S1 are processed. Twelve-dimensional feature parameters are extracted from the current signal, and vibration compensation is extracted from the vibration signal to provide feature input for subsequent nonlinear coupling analysis of current ripple torsional stress-strain. The specific operation is as follows.

[0065] Step S201: Perform a first-stage high-pass filter on the acquired raw current signal. The input is the digitized current signal output from step S1. The processing method uses a second-order Butterworth high-pass filter with a cutoff frequency set to 400Hz. The filter design considers group delay characteristics to reduce phase distortion of the filtered signal. The output is the current signal after the fundamental component has been attenuated. The judgment criterion is that frequency components below 400Hz are suppressed, while frequency components at 400Hz and above are retained.

[0066] Step S202: Perform a second-stage adaptive notch filter on the current signal after the first-stage high-pass filtering. The input is the current signal output from step S201. The processing method uses a lattice adaptive filter structure, with a center frequency set to 50Hz, a bandwidth set to 200Hz, and harmonic suppression covering integer multiples of 50Hz. The notch band avoids the 1kHz ripple fundamental frequency band. The filter coefficients are updated according to the minimum mean square criterion of the error signal, so that the notch center frequency follows the change of the power grid frequency, thereby automatically adjusting the notch center frequency according to the power grid frequency drift, and the notch depth is greater than 60dB. The output is the current signal after suppressing the 50Hz power frequency and its harmonic interference. The judgment criterion is that the amplitude attenuation of the 50Hz and its harmonic components is greater than 60dB.

[0067] Step S203 involves performing a third-stage five-point cubic smoothing on the current signal after the second-stage adaptive notch filtering. The input is the current signal output from step S202. The processing method is based on the cubic spline interpolation principle, performing cubic polynomial fitting on five consecutive sampling points to reduce high-frequency white noise. The output is a current signal with suppressed high-frequency random noise and minimal signal distortion. After three stages of preprocessing, the signal-to-noise ratio of the current signal is improved to over 35 dB.

[0068] S201 to S203 constitute a three-stage preprocessing flow for the current signal. The first stage suppresses the low-frequency fundamental component, the second stage eliminates 50Hz power frequency and its harmonic interference, and the third stage suppresses high-frequency white noise. After the above process, the signal-to-noise ratio in the current signal is improved to over 35dB, allowing the weak ripple changes related to the torsional state of the sunroof guide rail to be preserved, providing a stable signal basis for subsequent feature calculations.

[0069] Step S204: Extract 12-dimensional feature parameters from the current signal after three levels of preprocessing. The input is the current signal after three levels of preprocessing output from step S203. The processing method involves calculating the fundamental amplitude of ripple, the second harmonic amplitude of ripple, the third harmonic amplitude of ripple, the total harmonic distortion rate, the standard deviation of ripple amplitude, the frequency offset of ripple, the phase angle of ripple, the peak value of harmonic power spectral density and its corresponding frequency, the slope of the amplitude envelope, the frequency stability index, the amplitude variation characteristics, and the frequency component characteristics.

[0070] The fundamental ripple amplitude is taken as the amplitude of the 1kHz frequency component of the fast Fourier transform of the current signal after three stages of preprocessing. The second harmonic ripple amplitude is taken as the amplitude of the 2kHz frequency component. The third harmonic ripple amplitude is taken as the amplitude of the 3kHz frequency component.

[0071] The total harmonic distortion rate is the ratio of the root mean square value of the amplitudes of the 2nd to 5th harmonics to the amplitude of the ripple fundamental frequency. The standard deviation of the ripple amplitude is the statistical standard deviation of the ripple amplitude sequence. The ripple frequency offset is the deviation of the ripple fundamental frequency from its nominal value of 1kHz. The ripple phase angle is the phase difference between the ripple fundamental component and the reference signal. The peak values ​​of the harmonic power spectral density and their corresponding frequencies are obtained by searching for local maxima of the power spectral density.

[0072] The amplitude envelope is obtained by performing a Hilbert transform on the current signal after three stages of preprocessing and then taking its modulus. The slope of the amplitude envelope is the slope of the linear fit of the amplitude envelope. The frequency stability index is the coefficient of variation of the ripple frequency sequence, which is the ratio of the standard deviation of the ripple frequency sequence to the mean of the ripple frequency sequence. The amplitude variation feature is the ratio of the fundamental amplitude of the ripple to the standard deviation of the ripple amplitude. The frequency component feature is the product of the ripple frequency offset and the frequency stability index. The output is a feature vector composed of 12-dimensional feature parameters, with dimensions 1 to 12 corresponding to the 12 feature parameters mentioned above.

[0073] In step S205, the vibration acceleration signal acquired by the vibration sensor undergoes the same three-stage preprocessing as in steps S201 to S203, after which vibration compensation characteristics are extracted. The input is the digitized vibration signal output from step S1. The processing method involves calculating the ratio of the energy at the ripple characteristic frequency to the total energy across the entire frequency band of the vibration signal power spectrum. The ripple characteristic frequency includes frequencies of 1kHz, 2kHz, and 3kHz. The output is the vibration compensation amount.

[0074] The physical significance of vibration compensation lies in the fact that the vibration acceleration signal of the drive motor bearing housing reflects the coupling effect of vibration interference in the mechanical transmission chain on the current ripple signal. During the test, the torsional motion of the sunroof guide rail is transmitted to the drive motor shaft system through the coupling. The mechanical vibration of the shaft system changes the air gap length of the motor, causing a change in air gap magnetic reluctance, which in turn changes the air gap magnetic flux density, ultimately reflected in the change of motor current ripple. This modulation effect is superimposed on the current ripple change caused by the torsional state of the guide rail. If not distinguished, it will affect the mapping accuracy. The vibration compensation amount is input into the mapping model as the 8th core feature parameter, enabling the model to automatically learn the coupling relationship between vibration interference and current ripple characteristics during training, thereby effectively suppressing the interference component caused by vibration during prediction. By introducing the vibration compensation amount, the prediction stability of the mapping model under vibration interference conditions is improved, avoiding misjudgments caused by mechanical vibration.

[0075] S204 and S205 respectively extract multi-dimensional features of current ripple and extract vibration coupling compensation. The current ripple feature reflects the change in the load torque of the drive motor after being modulated by the torsional state of the guide rail, while the vibration compensation reflects the coupling effect of the mechanical vibration of the bearing housing on the current ripple through the air gap magnetic field. Together, they constitute a feature set containing both current ripple information and vibration coupling information, reducing the risk of misjudgment caused by mechanical vibration interference.

[0076] At the end of the automotive sunroof guide rail production line, inspection station S2 is performed on a certain model of sunroof guide rail. The drive motor speed is 1000 r / min, the torsional load is 100% of the rated torque, and the test lasts for 30 seconds. Step S1 collects current and vibration signals at a sampling frequency of 20 kHz and a quantization bit depth of 16 bits. In S2, the current signal is filtered by a 400 Hz high-pass filter, a 50 Hz adaptive notch filter, and a five-point triple smoothing, achieving a signal-to-noise ratio of 37 dB. 12-dimensional feature parameters are extracted, including a ripple fundamental amplitude of 0.42 A, a ripple second harmonic amplitude of 0.18 A, a ripple third harmonic amplitude of 0.09 A, a total harmonic distortion rate of 12.3%, a ripple amplitude standard deviation of 0.035 A, a ripple frequency offset of 0.8 Hz, a frequency stability index of 0.021, an amplitude variation characteristic of 12.0, a frequency component characteristic of 0.017 Hz, and a vibration compensation of 0.15. The single feature extraction time is less than 0.5 seconds. These features serve as input for subsequent S3, enabling the mapping model to distinguish differences in the torsional resistance of the guide rails. Compared to schemes without vibration compensation, the misclassification rate caused by vibration interference is reduced by approximately 30%.

[0077] At this point, step S2 outputs 12-dimensional feature parameters and vibration compensation, which serve as feature inputs for the nonlinear coupling analysis of current ripple and torsional stress-strain. These outputs connect to subsequent core feature selection and torsional performance mapping model construction, enabling the torsional performance evaluation of the sunroof guide rail to no longer rely on destructive torsional fatigue tests, but instead obtain a reasonable feature basis through the joint analysis of current ripple and vibration characteristics.

[0078] Furthermore, step S3 in the embodiment of this application is described: nonlinear coupling analysis of current ripple and torsional stress-strain. Step S3 selects core feature parameters that are highly correlated with torsional stress-strain from the 12-dimensional feature parameters and vibration compensation amount output in step S2, and establishes a nonlinear mapping relationship between current ripple features and torsional stress-strain, providing a stress-strain estimation basis for the subsequent torsional performance mapping model. The specific execution process is as follows.

[0079] Step S301: The mutual information method is used to calculate the mutual information value between each characteristic parameter and torsional stress and strain, and the core characteristic parameters are selected according to the mutual information threshold.

[0080] The input consists of the feature parameters from the 12-dimensional feature vector output in step S2, and the corresponding torsional stress or strain data obtained through traditional torsional fatigue testing. The processing method is based on information entropy theory, calculating the mutual information value between each feature parameter and the torsional stress and strain. The mutual information value characterizes the degree of statistical dependence between two random variables. The formula for calculating the mutual information value is:

[0081] in, For feature parameters With torsional stress or strain The mutual information value between them, in bits. For the feature parameters, take any dimension from the 12-dimensional feature parameters output in step S2. Torsional stress or strain; For feature parameters With torsional stress or strain The joint probability density function, For feature parameters The marginal probability density function, Torsional stress or strain The marginal probability density function.

[0082] Joint probability density function and marginal probability density function , By using feature parameters and torsional stress or strain The sample data were divided into several intervals, and the frequency of samples appearing in each interval was counted. The probability was approximated by the frequency. The larger the mutual information value, the higher the statistical dependence between the characteristic parameters and torsional stress and strain.

[0083] The output is the mutual information value between each feature parameter and torsional stress-strain. A mutual information threshold of 0.85 is set. When the mutual information value between a feature parameter and torsional stress-strain is greater than 0.85, the feature parameter is selected into the core feature parameter set. The mutual information threshold of 0.85 is determined by iterating through multiple candidate thresholds in the range of 0.70 to 0.95 with a step size of 0.05, filtering the core feature parameters and training the subsequent coupled model, and finally selecting the threshold corresponding to the highest goodness of fit on the validation set. The selection criterion is that feature parameters with a mutual information value greater than 0.85 are selected, while feature parameters with a mutual information value less than or equal to 0.85 are not selected.

[0084] Step S302: Determine the 8-dimensional core feature vector. The input is the set of core feature parameters obtained in step S301. Based on the screening criterion of a mutual information threshold of 0.85, 8-dimensional core feature parameters, including the ripple fundamental amplitude, are selected from the 12-dimensional feature vector. ripple second harmonic amplitude value Total Harmonic Distortion ripple amplitude standard deviation Peak value of harmonic power spectral density Slope of amplitude envelope Frequency stability index and vibration compensation amount .

[0085] The subscripts of the aforementioned current ripple features correspond one-to-one with the sequence numbers of the 12-dimensional feature parameters recorded in step S204. The vibration compensation amount, extracted from the self-vibration signal in step S205, does not belong to the 12-dimensional current ripple features, therefore... This is to avoid overlap with the 12th-dimensional current ripple feature. The output is a core feature vector composed of 8 core feature parameters:

[0086] in, The core feature vector contains 8 feature parameters. The fundamental amplitude of the ripple wave. This is the second harmonic amplitude of the ripple. Total harmonic distortion (THD) The standard deviation of the ripple amplitude. The peak value of the harmonic power spectral density. The slope of the amplitude envelope. This is the frequency stability index. This is the vibration compensation amount.

[0087] Step S303: Construct a backpropagation neural network optimized by a genetic algorithm, and set the genetic algorithm parameters and fitness function. The input is the 8-dimensional core feature vector output in step S302. , as well as the torsional stress and strain estimates obtained through traditional torsional fatigue tests.

[0088] The processing method employs a three-layer fully connected feedforward neural network topology. The input layer has 8 nodes, corresponding to 8-dimensional core feature parameters; the hidden layer has 15 nodes, with the hidden layer neurons using the hyperbolic tangent activation function; the output layer has 2 nodes, corresponding to the torsional stress estimate and the strain estimate, respectively, with the output layer using a linear activation function.

[0089] The number of hidden layer nodes, 15, was determined by traversing and searching within the range of 5 to 25 with a step size of 2. The minimum root mean square error of the verification set was used as the selection criterion, and 15 was determined to be the optimal number of hidden layer nodes.

[0090] Each node in the input layer is fully connected to each node in the hidden layer, and each node in the hidden layer is fully connected to each node in the output layer. The weight matrix from the input layer to the hidden layer is shown. The dimension is 15×8, and the bias vector is... The dimension is 15×1; the weight matrix from the hidden layer to the output layer. The dimension is 2×15, and the bias vector is... The dimension is 2×1. This embodiment uses a fully connected feedforward neural network, which does not involve convolutional layers and kernels, therefore there is no kernel size or convolutional layer activation function.

[0091] Genetic algorithms are used to optimize the initial weights and thresholds of the neural network, preventing the standard backpropagation algorithm from getting stuck in local optima. The training data is constructed as follows: a sample set is formed by collecting 8-dimensional core feature parameters and their corresponding torsional stress and strain estimates from the same batch of sunroof guide rails under the same test conditions. This sample set is divided into a training set and a validation set. The training set is used to train the neural network, and the validation set is used to calculate the fitness function of the genetic algorithm and evaluate the training effect.

[0092] The training set contains at least 200 samples, and the validation set contains at least 50 samples. Data cleaning standards include removing missing values ​​and using the 3σ criterion to remove outliers exceeding the mean plus or minus three standard deviations. The labeling rules stipulate that torsional stress and strain estimates are obtained through traditional torsional fatigue tests.

[0093] The genetic algorithm uses a population size of 50, a crossover probability of 0.7, a mutation probability of 0.01, and a maximum number of iterations of 200 generations. Individuals are encoded using real numbers; each individual is a real-number vector, with its elements corresponding to the weight matrices from the input layer to the hidden layer. Hidden layer bias vector Weight matrix from hidden layer to output layer and output layer bias vector The elements are arranged in row priority order.

[0094] The fitness function of a genetic algorithm is defined as the root mean square error (RMSE) of the neural network on the validation set. The RMSE is the square root of the average of the squares of the differences between the actual and estimated values ​​of torsional stress or strain in the validation set samples. A smaller fitness function value indicates a smaller prediction error on the validation set and a higher individual fitness.

[0095] Step S304 involves iterative optimization using a genetic algorithm to obtain initial weights and threshold configurations. The inputs are the genetic algorithm parameters and fitness function set in step S303. The process involves iteratively optimizing individuals in the population using three genetic operators: selection, crossover, and mutation.

[0096] The selection operator chooses superior individuals to proceed to the next generation based on their fitness function value; the crossover operator exchanges genes between two parent individuals with a crossover probability of 0.7; and the mutation operator randomly perturbs the genes of an individual with a mutation probability of 0.01. After a maximum of 200 iterations, the initial weights and threshold configurations corresponding to the individual with the smallest fitness function value are output. The criterion for determining the optimal initial weights and threshold configurations is that the individual with the smallest fitness function value is the one with the smallest initial weight and threshold configuration.

[0097] Step S305: Train the neural network and construct the feature stress mapping function. The inputs are the optimal initial weights and threshold configurations obtained in step S304, and the 8-dimensional core feature vector output in step S302. The approach involves using the optimal initial weights and thresholds as the initial parameters for the backpropagation neural network, and then training the network using the backpropagation algorithm.

[0098] The training loss function uses mean squared error (MSE), which is the average of the squares of the differences between the actual and estimated values ​​of torsional stress or strain in the training set samples. The training convergence objective is to achieve an MSE less than 0.001 or reach the maximum training epochs of 1000. The neural network training employs the Levenberg-Marquardt algorithm with a learning rate of 0.001 and a momentum factor of 0.9. Early stopping is triggered when the error no longer decreases after 10 consecutive training epochs. The output is the trained feature stress mapping function. This function takes an 8-dimensional core feature vector as input and outputs the estimated torsional stress and strain.

[0099] Step S306: Evaluate the goodness of fit of the feature stress mapping function. The inputs are the feature stress mapping function trained in step S305 and the validation set samples. The processing method is to calculate the goodness of fit of the feature stress mapping function on the validation set.

[0100] The goodness of fit is measured using the coefficient of determination, which is the ratio of the regression sum of squares to the total sum of squares, or equivalently calculated by subtracting the ratio of the residual sum of squares to the total sum of squares from 1. It is a common statistical indicator used to evaluate the goodness of fit of a regression model, and the closer the value is to 1, the better the fit.

[0101] The output is the goodness of fit of the characteristic stress mapping function. The criterion is that if the goodness of fit of the characteristic stress mapping function is greater than or equal to 0.92, the characteristic stress mapping function is considered to be able to capture the nonlinear mapping relationship between current ripple characteristics and torsional stress-strain.

[0102] In summary, step S3 outputs the trained feature stress mapping function. The feature stress mapping function takes an 8-dimensional core feature vector as input and outputs the torsional stress estimate and strain estimate, and connects to the subsequent construction of the current ripple torsional performance mapping model.

[0103] Following this is step S4 in this embodiment of the application: the construction of the current ripple torsional performance mapping model. Based on the 8-dimensional core feature parameters selected in step S3, a quantitative mapping model between current ripple characteristics and material torsional performance indicators is constructed, and the torsional stiffness, torsional strength and fatigue life are output as three torsional performance evaluation results. This can be performed in the following manner.

[0104] Step S401: Construct a 3-layer backpropagation neural network, with the 8-dimensional core feature parameters output from step S3 as the input. It also includes three torsional performance indicators: torsional stiffness, torsional strength, and fatigue life.

[0105] The processing method involves determining the number of nodes in the input layer, the number of nodes in the hidden layer, the number of nodes in the output layer, the activation functions of each layer, and the connection relationships between each layer in a three-layer backpropagation neural network.

[0106] The number of nodes in the input layer is set to 8, corresponding to 8-dimensional core feature parameters. Each input node receives a 1-dimensional feature parameter value. The number of hidden layer nodes is set to 15, and the hidden layer neurons use the hyperbolic tangent activation function. The mathematical expression of the hyperbolic tangent function is:

[0107] in, The weighted sum of inputs to the hidden layer neurons is dimensionless. is a natural constant with a value of 2.71828; This is the output of the hyperbolic tangent activation function, with a value ranging from -1 to 1.

[0108] The output layer has 3 nodes, corresponding to the three torsional performance indicators: torsional stiffness, torsional strength, and fatigue life. The output layer uses a linear activation function, which is:

[0109] in, The weighted sum of the inputs and outputs of the output layer neurons; This represents the output value of the output layer neurons.

[0110] Each node in the input layer is fully connected to each node in the hidden layer, and each node in the hidden layer is fully connected to each node in the output layer. The weight matrix from the input layer to the hidden layer is shown. The dimension is 15×8, and the bias vector is... The dimension is 15×1; the weight matrix from the hidden layer to the output layer. The dimension is 3×15, and the bias vector is... The dimension is 3×1.

[0111] The current ripple anti-torsion performance mapping model adopts a fully connected feedforward neural network, which does not involve convolutional layers and kernels. Therefore, there are no kernel size and convolutional layer activation functions. The output is a 3-layer backpropagation neural network structure.

[0112] Step S402: Perform z-score normalization on the input feature parameters, with the input being the 8-dimensional core feature parameters output from step S3. The processing method involves processing the 8-dimensional core feature parameters. Each 1-dimensional feature parameter in the model is independently standardized using z-score.

[0113] For the Dimensional feature parameters, calculate the th feature parameter The mean of the 3D feature parameters on the training set and standard deviation Then the first Subtract the mean from each value of the feature parameter. Then divide by the standard deviation This yields standardized variables. The eight feature parameters are processed independently, with no cross-coupling between dimensions. The output consists of eight standardized feature parameters, each with a mean of 0 and a standard deviation of 1.

[0114] Step S403: Determine the number of hidden layer nodes. The input is the candidate range of the number of hidden layer nodes, which is 5 to 25, and the step size is 2. The candidate number of nodes is 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25.

[0115] The approach involves constructing a three-layer backpropagation neural network for each candidate node, training it using the same training and validation sets, and calculating the root mean square error (RMSE) of the neural network on the validation set for each candidate node. The RMSE is the square root of the average of the squares of the differences between the baseline torsional resistance values ​​of the validation set samples and the neural network output values.

[0116] The output is the number of candidate nodes with the minimum root mean square error of the validation set. The criterion is that the number of candidate nodes with the minimum root mean square error of the validation set is the optimal number of hidden layer nodes, and 15 was ultimately determined to be the optimal number of hidden layer nodes.

[0117] Step S404: Set the neural network training parameters and loss function. The input is the 3-layer backpropagation neural network constructed in step S401 and the optimal number of hidden layer nodes (15) determined in step S403.

[0118] The approach involves setting the learning rate of the neural network to 0.001 and the momentum factor to 0.9. The Levenberg-Marquardt algorithm is used for training. The loss function is the mean squared error, which is the average of the squared differences between the baseline torsional resistance values ​​of the training set samples and the neural network output values. The training convergence objective is for the mean squared error to be less than 0.001 or to reach the maximum training epochs of 1000. An early stopping mechanism is triggered when the training error no longer decreases after 10 consecutive epochs.

[0119] The output contains the neural network training parameters and loss function configuration.

[0120] Step S405: Construct a training sample set and torsional performance benchmark values. The inputs are the current ripple characteristic parameters collected from the same batch of sunroof guide rails under the same test conditions, and the torsional performance benchmark values ​​obtained through traditional torsional fatigue tests.

[0121] The processing method involves creating a sample by collecting 8-dimensional core feature parameters and their corresponding torsional performance benchmark values ​​from the same batch of sunroof guide rails under the same test conditions, with a training sample size of no less than 200 groups. The data cleaning standard is to remove missing values ​​and use the 3σ criterion to remove outlier samples that exceed the mean plus or minus 3 times the standard deviation.

[0122] The labeling rules stipulate that the baseline value for torsional performance is obtained through traditional torsional fatigue testing, including the slope of the torque-rotation curve of torsional stiffness, the maximum torsional strength, and the standard fatigue life hours. The fatigue life baseline value is measured in hours and is consistent with the fatigue life estimate y3 output in step S504 in terms of dimensions and units. The output is a training sample set.

[0123] Step S406: Train and validate the mapping model. The inputs are the training sample set constructed in step S405, the 3-layer backpropagation neural network constructed in step S401, the optimal number of hidden layer nodes (15) determined in step S403, and the training parameters and loss function set in step S404.

[0124] The approach employed is to train the neural network using leave-one-out cross-validation. Leave-one-out cross-validation involves selecting one set of samples from the training set as validation samples each time, using the remaining samples as training samples, and repeating this process until all samples have been used as validation samples. After model training, generalization performance is validated using an independent test set with at least 50 samples, and the model's generalization error is less than or equal to 5%.

[0125] The output is the trained current ripple torsional resistance performance mapping model. The criterion is that the model's generalization error is less than or equal to 5%, indicating that the current ripple torsional resistance performance mapping model meets the generalization performance requirements.

[0126] Therefore, step S4 outputs the trained current ripple torsional resistance performance mapping model. The current ripple torsional resistance performance mapping model uses 8-dimensional core feature parameters. As input, and with torsional stiffness, torsional strength, and fatigue life as output, it provides a model reasoning basis for the torsional performance evaluation in the subsequent step S5.

[0127] Finally, step S5 in the embodiment of this application, the torsional performance evaluation, is described. First, a torsional test is performed on the sunroof guide rail to be tested. Then, the current ripple characteristic parameters are extracted and input into the trained current ripple torsional performance mapping model. The evaluation results of torsional stiffness, torsional strength and fatigue life are output. Finally, the torsional performance evaluation process is completed. The specific implementation process is as follows.

[0128] Step S501: Perform a torsion test on the sunroof guide rail under test. The inputs are the sunroof guide rail under test, the drive motor, and the torque sensor. The process involves mounting the sunroof guide rail under test on a dedicated torsion test bench, and connecting the drive motor to the sunroof guide rail via a coupling.

[0129] The drive motor speed setting is determined by multiplying the rated operating speed by a speed coefficient, which is between 0.5 and 0.8. The specific value is determined based on the model of the sunroof guide rail under test and the test requirements. The speed setting is controlled within the range of 500 r / min to 2000 r / min. The torsional load setting is determined by multiplying the rated torque by a load coefficient, which is between 0.8 and 1.2.

[0130] The specific values ​​are determined based on the model of the sunroof guide rail under test and the test requirements. The applied torque value is monitored in real time using a torque sensor. When the deviation between the monitored value and the torsional load setpoint exceeds 5%, the drive motor output is adjusted to maintain the torsional load near the setpoint. The test duration is no less than 30 seconds to obtain sufficient steady-state data for feature extraction.

[0131] The output is the drive motor current signal of the sunroof guide rail under controlled torsion conditions. The judgment criteria are that the drive motor speed is within the range of 500 r / min to 2000 r / min, the torsional load is within the range of 80% to 120% of the rated torque, and the test duration is not less than 30 seconds.

[0132] Step S502: Extract 12-dimensional feature parameters and 8-dimensional core feature parameters from the drive motor current signal of the sunroof guide rail under test. The input is the drive motor current signal acquired in step S501.

[0133] The processing method employs the same preprocessing and feature calculation methods as in step S2, extracting 12-dimensional feature parameters from the drive motor current signal of the sunroof guide rail under test. These 12-dimensional feature parameters include the ripple fundamental amplitude, ripple second harmonic amplitude, ripple third harmonic amplitude, total harmonic distortion rate, ripple amplitude standard deviation, ripple frequency offset, ripple phase angle, peak value of harmonic power spectral density and its corresponding frequency, amplitude envelope slope, frequency stability index, amplitude variation characteristics, and frequency component characteristics.

[0134] Then, based on the screening results of step S3, eight core feature parameters are extracted from the 12-dimensional feature parameters. These eight core feature parameters include the fundamental amplitude of the ripple, the second harmonic amplitude of the ripple, the total harmonic distortion rate, the standard deviation of the ripple amplitude, the peak value of the harmonic power spectral density, the slope of the amplitude envelope, the frequency stability index, and the vibration compensation amount. The vibration compensation amount is extracted from the vibration acceleration signal collected by the vibration sensor in step S2 after undergoing the same three-stage preprocessing as in steps S201 to S203.

[0135] The output is the 8-dimensional core feature parameters of the sunroof guide rail under test. The judgment criterion is that the 8-dimensional core feature parameters are consistent with the 8-dimensional core feature parameters determined in step S3.

[0136] Step S503: Input the 8-dimensional core feature parameters into the trained current ripple torsional resistance mapping model and perform forward propagation calculation. The input is the 8-dimensional core feature parameters output in step S502. 8-dimensional core feature parameters Let it be an 8-dimensional vector, denoted as .

[0137] in, The fundamental amplitude of the ripple wave. This is the second harmonic amplitude of the ripple. Total harmonic distortion (THD) The standard deviation of the ripple amplitude. The peak value of the harmonic power spectral density. The slope of the amplitude envelope. This is the frequency stability index. This is the vibration compensation amount. The processing method involves using the 8-dimensional core feature parameters. Perform z-score standardization.

[0138] For the 3D feature parameters According to the training result obtained in step S4, the first Mean of feature parameters and standard deviation Standardization is performed to obtain the first 3D standardized feature parameters Vibration compensation amount The corresponding components are obtained through the same standardization. Each of the 8 feature parameters is standardized independently, with no cross-coupling between dimensions. The standardized result is an 8-dimensional standardized feature vector. .

[0139] Standardize the feature vector Input the trained current ripple torsional resistance performance mapping model. The current ripple torsional resistance performance mapping model adopts the 3-layer backpropagation neural network trained in step S4. The network structure has 8 nodes in the input layer, 15 nodes in the hidden layer, and 3 nodes in the output layer.

[0140] Weight matrix from input layer to hidden layer The dimension is 15×8, and the hidden layer bias vector is... The dimension is 15×1; the weight matrix from the hidden layer to the output layer. The dimension is 3×15, and the output layer bias vector is... The dimension is 3×1. The hidden layer neurons use the hyperbolic tangent activation function, and the output layer uses the linear activation function.

[0141] Hidden layer Weighted sum of neurons Calculate according to the following formula:

[0142] in, For the hidden layer The weighted sum of 10 neurons, dimensionless; Number the neurons in the hidden layer. ; For the input layer The node is connected to the hidden layer. The weight of each node is dimensionless. For the first Dimensionally standardized characteristic parameters, dimensionless; For the hidden layer The bias of each neuron is dimensionless.

[0143] Hidden layer The output of each neuron Calculate according to the following formula:

[0144] in, For the hidden layer The output of each neuron is dimensionless. For the hidden layer The weighted sum of neurons; It is a natural constant with a value of 2.71828.

[0145] Output layer Weighted sum of neurons The weighted sum is calculated based on the output layer neurons. The output layer uses a linear activation function. The output of each neuron Equal to the output layer Weighted sum of neurons The output layer node 1 corresponds to the torsional stiffness evaluation value, node 2 corresponds to the torsional strength evaluation value, and node 3 corresponds to the fatigue life evaluation value.

[0146] The output consists of three output values ​​from the output layer of the current ripple torsional resistance performance mapping model. , , These correspond to the evaluation results of torsional stiffness, torsional strength, and fatigue life, respectively. The judgment criterion is that the number of nodes in the output layer is 3, corresponding to torsional stiffness, torsional strength, and fatigue life, respectively.

[0147] Step S504: Output the torsional performance evaluation results in a structured format, with the input being the three output values ​​from step S503. , , The processing method involves presenting the three output values ​​in a structured format.

[0148] The torsional stiffness output is a percentage of the standard value. The standard value is defined as the average torsional stiffness of qualified sunroof guide rails in the same batch, denoted as . Torsional stiffness output value Calculate according to the following formula:

[0149] in, Output percentage of torsional stiffness; This is the torsional stiffness evaluation value output by node 1 of the output layer; This represents the average torsional stiffness of qualified sunroof guide rails from the same batch. The output range is 70% to 130%. The judgment criterion is the percentage of torsional stiffness output. The range is between 70% and 130%.

[0150] The torsional strength rating is output in four levels: Excellent, Good, Medium, and Poor. The evaluation criteria are determined based on a percentage threshold of the rated torque. (Torsional strength percentage) Calculate according to the following formula:

[0151] in, The percentage is the torsional strength. , This is the torsional strength evaluation value output by node 2 of the output layer. This represents the rated torque of the sunroof guide rail under test. Each grade uses mutually exclusive and interconnected intervals: the superior grade corresponds to... Good grade correspondence Intermediate level correspondence Difference correspondence The criterion is the percentage of torsional strength. Output the corresponding level when it falls into the unique range mentioned above.

[0152] The fatigue life estimate is output in hours, denoted as The prediction range is 0 to 10,000 hours. The criterion is the estimated fatigue life. Within the range of 0 to 10000 hours, the output is the torsional performance evaluation result of the sunroof guide rail under test.

[0153] At the final inspection station of the automotive sunroof guide rail production line, steps S1 to S5 are performed on a certain model of sunroof guide rail. The drive motor speed is set to 1000 r / min, the torsional load is applied to 100% of the rated torque, and the test duration is 30 seconds. Step S1 collects current ripple signals and vibration signals; step S2 extracts 12-dimensional feature parameters and vibration compensation; step S3 filters 8-dimensional core feature vectors using the mutual information method; step S4 trains the current ripple torsional performance mapping model; and step S5 inputs the 8-dimensional core feature vectors into the trained current ripple torsional performance mapping model to output torsional stiffness, torsional strength, and fatigue life.

[0154] The input data format is an 8-dimensional core feature parameter vector. The input current ripple torsional resistance performance mapping model is then normalized using z-score. The current ripple torsional resistance performance mapping model outputs three values. , , The output data generation rules are as follows: according to Generate torsional stiffness percentage; according to Generate torsional strength percentage and torsional strength grade according to threshold range; It is directly used as the fatigue life estimate output.

[0155] The results were verified on 200 test samples. The torsional stiffness assessment error was less than or equal to 4.2%, the accuracy rate of torsional strength grade determination was 96.5%, the fatigue life prediction error was less than or equal to 5.1%, and the single-piece testing time was less than or equal to 45 seconds.

[0156] Finally, in summary, step S5 outputs the torsional stiffness percentage, torsional strength grade, and fatigue life estimate of the sunroof guide rail under test, completing the torsional performance evaluation process. The evaluation results can be stored in the testing database to provide historical testing data for subsequent residual life prediction.

[0157] Furthermore, embodiments of this application also provide detection database management and residual lifetime prediction functions.

[0158] The testing database contains no fewer than 5,000 historical testing records. Each record includes data fields such as rail model, production batch, test date, current ripple characteristic parameters, evaluation results, and re-inspection conclusions. The rail model field records the product model code of the rail. The production batch field records the manufacturing batch identifier of the rail.

[0159] The Test Date field records the specific date and time of the test execution. The Current Ripple Characteristic Parameter field stores 12-dimensional raw characteristic parameters and 8-dimensional core characteristic parameters in the form of a structured array. The Evaluation Result field stores the quantitative evaluation values ​​of three torsional performance indicators. The Re-inspection Conclusion field records the verification results of subsequent re-inspections.

[0160] Based on historical testing data in the testing database, time series analysis is used to predict the residual service life of the guide rail. The time series analysis method employs an autoregressive integral moving average model. The autoregressive integral moving average model consists of an autoregressive component, a differencing component, and a moving average component; the model order is denoted as [model order denoted as 1]. , , .in, Let the order be the autoregressive order. Let be the difference order. This represents the order of the moving average.

[0161] , , The value of is determined by traversing the candidate order array, merging and comparing it according to the Akaike Information Criterion. In the candidate order combinations, Take integers from 0 to 5. Take an integer from 0 to 2. Take integers from 0 to 5. The Akaike Information Criterion is calculated according to the following formula:

[0162] in, The value of the Akaike Information Criterion is dimensionless. The number of parameters in the autoregressive integral moving average model is dimensionless. Let be the dimensionless likelihood function value of the autoregressive integral moving average model. The order of the autoregressive integral moving average model is determined by the combination of orders that minimizes the Akaike information criterion value.

[0163] The autoregressive integral moving average model takes as input the torsional performance assessment results from the five most recent consecutive tests and outputs the predicted residual service life. The model parameters are fitted using the maximum likelihood estimation method. The maximum likelihood estimation method aims to maximize the probability of occurrence of the torsional performance assessment results in the historical test data to solve for the model parameters.

[0164] The lead time for predicting residual useful life is 1000 to 5000 hours. The specific lead time is adaptively adjusted based on the deviation between the current assessment results and the warning threshold. The deviation is calculated using the following formula:

[0165] in, The degree of deviation is dimensionless; This is the current assessment result; This is the warning threshold. When When the value is greater than 0 and less than or equal to 0.1, the forecast lead time is set to 3000h to 5000h. When the value is greater than 0.1 and less than or equal to 0.2, the forecast lead time is set to 1000h to 2000h. When the value is greater than 0.2, the forecast lead time is set to 1000h.

[0166] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0167] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for evaluating the torsional performance of sunroof guide rails based on current ripple and vibration characteristics, comprising: On a special torsion test bench for sunroof guide rails, a drive motor is connected to the sunroof guide rail via a coupling, and the drive motor outputs periodic torsion motion; the feature is that: the drive motor current ripple signal and the drive motor bearing housing vibration signal are acquired synchronously, and the drive motor current ripple signal and the drive motor bearing housing vibration signal are digitized through a multi-channel data acquisition card; The digitized drive motor current ripple signal is subjected to the first-level high-pass filtering, the second-level adaptive notch filtering, and the third-level five-point triple smoothing process in sequence to extract 12-dimensional feature parameters and extract vibration compensation amount from the drive motor bearing housing vibration signal. The mutual information method is used to calculate the mutual information value between each feature parameter in the 12-dimensional feature parameters and the torsional stress and strain. The core feature parameters are selected according to the mutual information threshold. A feature stress mapping function is constructed with an 8-dimensional core feature vector as input and the torsional stress estimate and strain estimate as output. The feature stress mapping function is implemented by a backpropagation neural network optimized by a genetic algorithm. The eight core feature parameters of the sunroof guide rail under test are input into the trained current ripple torsional performance mapping model, and the torsional stiffness evaluation value, torsional strength evaluation value, and fatigue life evaluation value are output.

2. The method according to claim 1, characterized in that: The drive motor current ripple signal is acquired by a Hall effect closed-loop current sensor, which is installed on any one phase of the three-phase power line of the drive motor using a through-hole mounting method. The vibration signal of the drive motor bearing housing is collected by a piezoelectric accelerometer, which is fixed to the end face of the drive motor bearing housing by a threaded connection. The acquisition trigger terminals of the Hall effect closed-loop current sensor and the piezoelectric accelerometer are connected to the same trigger source, so that the current ripple signal of the drive motor and the vibration signal of the drive motor bearing housing are kept in sync in terms of timing, and the timestamps are aligned with each other.

3. The method according to claim 1, characterized in that: The digitized drive motor current ripple signal is subjected to the first-stage high-pass filtering, the second-stage adaptive notch filtering, and the third-stage five-point triple smoothing process in sequence, including: using a second-order Butterworth high-pass filter for the first-stage high-pass filtering, with the cutoff frequency set to 400Hz; The second-stage adaptive notch filter is performed using a lattice adaptive filter structure, with the center frequency set to 50Hz and the bandwidth set to 200Hz. The filter coefficients are updated according to the minimum mean square criterion of the error signal, so that the notch center frequency follows the change of the power grid frequency. Based on the principle of cubic spline interpolation, cubic polynomial fitting is performed on 5 consecutive sampling points, followed by a third-level five-point cubic smoothing.

4. The method according to claim 1, characterized in that: Extracting vibration compensation from the vibration signal of the drive motor bearing housing includes: calculating the ratio of the energy of the power spectrum of the drive motor bearing housing vibration signal at the ripple characteristic frequency to the energy of the entire frequency band, where the ripple characteristic frequency includes 1kHz, 2kHz and 3kHz. The vibration compensation amount is extracted after undergoing the same first-stage high-pass filtering, second-stage adaptive notch filtering, and third-stage five-point triple smoothing process as the drive motor current ripple signal. Mechanical vibration of the shaft system changes the air gap length of the motor, causing changes in air gap magnetic reluctance, which in turn changes the air gap magnetic flux density, and consequently, the current ripple signal of the drive motor changes accordingly.

5. The method according to claim 1, characterized in that: The mutual information method was used to calculate the mutual information values ​​between each feature parameter in the 12-dimensional feature parameters and the torsional stress and strain, and the core feature parameters were selected based on the mutual information threshold, including: Calculate the mutual information value between each characteristic parameter in the 12-dimensional characteristic parameters and the torsional stress and strain. The formula for calculating the mutual information value is as follows: in, For feature parameters With torsional stress or strain The mutual information value between them, in bits. For the feature parameters, any dimension from the 12-dimensional feature parameters can be selected. Torsional stress or strain; For feature parameters With torsional stress or strain The joint probability density function, For feature parameters The marginal probability density function, Torsional stress or strain The marginal probability density function; The mutual information threshold is set to 0.85, and feature parameters with mutual information values ​​greater than 0.85 are selected into the core feature parameter set. Eight core feature parameters were selected from the 12 feature parameters. These eight core feature parameters are: ripple fundamental amplitude, ripple second harmonic amplitude, total harmonic distortion rate, ripple amplitude standard deviation, peak value of harmonic power spectral density, slope of amplitude envelope, frequency stability index, and vibration compensation.

6. The method according to claim 1, characterized in that: In the backpropagation neural network optimized by the genetic algorithm: the population size of the genetic algorithm is set to 50, the crossover probability is set to 0.7, the mutation probability is set to 0.01, the maximum number of iterations is set to 200 generations, the individual encoding method is real number encoding, and the fitness function is the root mean square error of the backpropagation neural network on the validation set. The backpropagation neural network adopts a three-layer fully connected feedforward neural network topology, with 8 nodes in the input layer, 15 nodes in the hidden layer, and the hyperbolic tangent activation function for the hidden layer neurons. The output layer has 2 nodes and uses a linear activation function. The goodness of fit of the characteristic stress mapping function on the validation set is greater than or equal to 0.

92.

7. The method according to claim 1, characterized in that: The current ripple anti-torsion performance mapping model adopts a three-layer backpropagation neural network with 8 nodes in the input layer, 15 nodes in the hidden layer, hyperbolic tangent activation function for the hidden layer neurons, 3 nodes in the output layer, and linear activation function for the output layer. The weight matrix W1 from the input layer to the hidden layer has a dimension of 15×8, and the weight matrix W2 from the hidden layer to the output layer has a dimension of 3×15. Each of the 1-dimensional feature parameters in the 8-dimensional core feature parameters is independently standardized by z-score, and there is no cross-coupling between the dimensions.

8. The method according to claim 1, characterized in that: When tested on a special torsion test bench for sunroof guide rails, the speed of the drive motor and the torsional load remained stable, and the motor itself remained basically constant. The change in current ripple characteristics in the drive motor current ripple signal originates from the mechanical response of the tested sunroof guide rail under torsional load, which modulates the motor load torque through the transmission chain. The mechanical response includes local stiffness changes caused by the evolution of microscopic defects in the material. The torsional stress and strain estimates were calibrated and verified using stress and strain data obtained from traditional torsional fatigue tests.

9. The method according to claim 1, characterized in that: Outputs torsional stiffness assessment value, torsional strength assessment value, and fatigue life assessment value, including: the torsional stiffness assessment value is output as a percentage of the average torsional stiffness of the same batch of qualified sunroof guide rails, with an output range of 70% to 130%; The torsional strength assessment value is output in four grades: excellent, good, medium, and poor. Excellent corresponds to a torsional strength percentage greater than or equal to 95%, good corresponds to a torsional strength percentage greater than or equal to 90% and less than 95%, medium corresponds to a torsional strength percentage greater than or equal to 85% and less than 90%, and poor corresponds to a torsional strength percentage less than 85%. The fatigue life assessment value is output in hours, with a prediction range of 0 to 10,000 hours. A testing database containing no less than 5,000 historical testing records was established. An autoregressive integral moving average model was used to predict the residual service life of the guide rail. The input of the autoregressive integral moving average model was the torsional performance evaluation results of the five most recent consecutive tests, and the output was the predicted value of the residual service life. The prediction lead was 1,000 hours to 5,000 hours.

10. The method according to claim 1, characterized in that: The training set of the current ripple torsional resistance mapping model shall have no less than 200 sets of samples and no less than 50 sets of samples. The neural network shall be trained using leave-one cross-validation. After the current ripple torsional resistance performance mapping model is trained, its generalization performance is verified using an independent test set with no less than 50 test set samples. The generalization error of the current ripple torsional resistance performance mapping model is less than or equal to 5%.

Images