A method for monitoring material wear based on shear wave active ultrasound and passive acoustic emission

Abstract

This invention proposes a method for co-monitoring material wear based on active ultrasonic and passive acoustic emission using shear waves. The method involves synchronously acquiring active and passive shear wave acoustic emission signals using a combined active and passive shear wave sensor. The active and passive shear wave acoustic emission signals are preprocessed to analyze the modulation effect of the active shear wave echo signal on the passive shear wave acoustic emission signal. Adaptive filtering and neural network regression algorithms are used to reconstruct the passive shear wave acoustic emission signal, and the optimal reconstructed passive shear wave acoustic emission signal is selected. The separated wear components are obtained through blind source separation. Data-level fusion and feature-level fusion are performed. The data-level fusion feature vector and the feature-level fusion feature vector are used for co-representation. A wear depth regression network outputs the estimated wear depth and surface morphology feature index under the current working condition, and a wear type classifier outputs the wear mechanism category. This achieves real-time, high-precision co-representation of wear depth, surface morphology features, and wear mechanism type.

Description

Technical Field

[0001] This invention relates to the technical field of ultrasonic nondestructive testing, and more particularly to a method for monitoring material wear based on transverse waves. Background Technology

[0002] Acoustic emission (AE) technology can capture in real time the sudden signals generated by transient processes such as microcrack propagation, plastic deformation, or oil film rupture during the friction and wear of materials. It is currently widely used in the fields of mechanical condition monitoring and fault diagnosis. At the same time, ultrasonic testing, as a quantitative measurement method, has significant advantages over optical methods, such as small sensor size and convenient installation, and is particularly suitable for real-time, in-situ measurement in confined spaces such as inside an engine.

[0003] Currently, two independent wear monitoring technologies exist in the industry: passive acoustic emission (AE) monitoring and active ultrasonic pulse-echo detection. Passive acoustic emission relies on capturing the elastic waves spontaneously released during transient wear damage to qualitatively identify the timing of wear occurrence and distinguish different wear mechanisms. Active ultrasonic detection, by actively emitting elastic waves and receiving reflected echoes from the interface, can quantitatively measure the amount of wear removed from the friction interface. The two technologies are naturally complementary. Theoretically, using a combined active and passive acoustic emission monitoring approach can simultaneously obtain quantitative wear dimensions and damage mechanism information. However, current technologies have not yet established a mature combined active and passive monitoring architecture, and several core technological bottlenecks remain unresolved.

[0004] First, the vibration modulation effect of active ultrasonic high-frequency disturbance on passive shear wave acoustic emission signal leads to distortion of passive AE signal, thus affecting the accuracy of wear mechanism discrimination. Second, the decoupling and separation of different mechanism components in active and passive signals under complex wear mechanisms (such as when abrasive wear and adhesive wear coexist). Third, how to effectively integrate the wear depth information obtained by active shear wave and the wear mechanism information obtained by passive shear wave AE to achieve synergistic characterization of wear depth, surface morphology and wear mechanism. Summary of the Invention

[0005] To address the technical challenges of existing technologies, such as the difficulty in achieving vibration modulation interference of active ultrasonic pulse echoes on passive acoustic emission (AE) signals, signal decoupling and separation under complex wear mechanisms, and multi-source information fusion of wear depth, this invention proposes a method for co-monitoring material wear based on transverse waves using active ultrasonic and passive acoustic emission. This method establishes a measurement architecture that coordinates active transverse wave ultrasonic detection and passive transverse wave acoustic emission, utilizing both active and passive transverse wave piezoelectric sensors to simultaneously perform active excitation and passive reception tasks. Based on this, the acquired active echo and passive AE signals are sequentially processed for signal decoupling and restoration, multi-dimensional feature fusion, and wear depth regression modeling. Ultimately, this achieves real-time, high-precision co-characterization of wear depth, surface morphology features, and wear mechanism types.

[0006] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0007] A method for co-monitoring material wear based on transverse wave-driven active ultrasound and passive acoustic emission includes:

[0008] S1: Using the same clock reference, the active shear wave echo signal and the passive shear wave acoustic emission signal are synchronously acquired through the active and passive shear wave co-sensor.

[0009] S2: Preprocess the active shear wave echo signal and the passive shear wave acoustic emission signal, and analyze the modulation effect of the active shear wave echo signal on the passive shear wave acoustic emission signal.

[0010] S3: The passive transverse wave acoustic emission signal is restored by using adaptive filtering and neural network regression algorithms respectively, and the optimal restored passive transverse wave acoustic emission signal is selected.

[0011] S4: Using the restored passive transverse wave acoustic emission signal and the preprocessed active transverse wave echo signal as joint inputs, the separated wear component is obtained through blind source separation.

[0012] S5: Data-level fusion is performed using the restored passive transverse wave acoustic emission signal and the preprocessed active transverse wave echo signal as joint inputs, and feature-level fusion is performed using the separated wear component as input.

[0013] S6: Using the multi-dimensional fused feature vector obtained by the collaborative representation of data-level fused feature vector and feature-level fused feature vector as input, the wear depth regression network outputs the wear depth estimate and surface morphology feature index under the current working condition, and the wear mechanism category is output through the wear type classifier.

[0014] Furthermore, the modulation effect of the active shear wave echo signal on the passive shear wave acoustic emission signal is analyzed, including:

[0015] A1. Based on the preprocessed active shear wave echo signal The amplitude attenuation characteristics and flight time variations are used to extract parameters characterizing the interface contact stiffness changes within the sensor coverage area.

[0016] ;

[0017] in, The reference amplitude under no-wear condition is, This represents the echo amplitude of the preprocessed active shear wave echo signal.

[0018] A2. Based on Hertzian contact theory, establish a mapping function F between interface stiffness variation and modulation factor, and characterize the interface contact stiffness variation as a parameter. Mapped to modulation factor:

[0019] ;

[0020] A3. Modulation factor Perform spline interpolation to align the preprocessed passive transverse wave acoustic emission signal. Sampling rate, to obtain a continuous modulation factor sequence .

[0021] Furthermore, signal restoration based on adaptive filtering is employed, including: restoring the passive transverse wave acoustic emission signal using adaptive transverse filtering (FIR) and LMS algorithms.

[0022] B1. The preprocessed active shear wave echo signal Dynamic modulation is performed using a continuous modulation factor c[n] to obtain the dynamically modulated active shear wave echo signal:

[0023] ;

[0024] B2. The dynamically modulated active shear wave echo signal As a reference input, the passive transverse wave acoustic emission signal As the desired response, for each sampling point n of the passive transverse wave acoustic emission signal, an adaptive transverse filtering FIR iterative calculation based on the LMS algorithm is performed:

[0025] B21. Calculate the estimated modulation interference components:

[0026] ;

[0027] in, ; The reference input vector is of order L; This represents the convolution operation.

[0028] B22. Calculate the error signal, i.e., the passive transverse wave acoustic emission signal reconstructed from the current sampling point:

[0029] ;

[0030] B23. Update the weight vector according to the LMS (Least Mean Square) criterion:

[0031] ;

[0032] ;

[0033] in, δ is the step size factor, α is the convergence coefficient used to compromise between convergence speed and steady-state error; δ is the regularization minimum constant used to prevent division by zero.

[0034] B3. Select the error signal after the final iteration as the first passive transverse wave acoustic emission signal for signal restoration based on adaptive filtering. .

[0035] Furthermore, signal restoration based on neural network regression is employed, including: using dynamically modulated active shear wave echo signals. and preprocessed passive transverse wave acoustic emission signal The synchronous timing segment constructs the input dual-channel timing signal Based on the first temporal convolutional network TCN, dual-channel time-series signals The signal was processed to obtain the second passive transverse wave acoustic emission signal based on neural network regression. The first temporal convolutional network (TCN) adds a subtraction layer to the original TCN network to process the preprocessed passive transverse wave acoustic emission signal. Modulation interference components estimated by the original TCN network By subtracting point by point, the second passive transverse wave acoustic emission signal based on neural network regression is obtained. .

[0036] Furthermore, the optimal restored passive transverse wave acoustic emission signal is selected, including:

[0037] C1. Using the preprocessed active shear wave echo signal The passive transverse wave acoustic emission signal after the two paths are restored, serving as a reference signal. Respectively with reference signal To compare the two paths, the amplitude squared coherence coefficients between the restored passive transverse wave acoustic emission signals and the reference signal were calculated:

[0038] ;

[0039] In the formula, The cross-power spectral density between the restored passive signal and the active shear wave signal. and These are the autopower spectral densities of the restored passive signal and the reference signal, respectively.

[0040] C2. Calculate the average coherence coefficients of the passive shear wave acoustic emission signals and the reference signal after restoration for both paths within the effective frequency band of the passive shear wave acoustic emission sensor:

[0041] ;

[0042] In the formula, As the lower limit of the effective frequency band, f H This is the upper limit of the effective frequency band;

[0043] C3. Compare the average coherence coefficients of the two restoration paths, and select the restoration signal corresponding to the one with the smaller average coherence coefficient as the final restored passive transverse wave acoustic emission signal. :

[0044] .

[0045] Furthermore, the separated wear component is obtained through blind source separation, including:

[0046] D1. To restore the passive transverse wave acoustic emission signal and preprocessed active shear wave echo signal For joint input The FastICA algorithm is used for joint blind source separation to decouple the signal components corresponding to abrasive wear and adhesive wear, thus obtaining the first abrasive wear component. With the first adhesive wear component ;

[0047] D2. To restore the passive transverse wave acoustic emission signal and preprocessed active shear wave echo signal For joint input, a second temporal convolutional network (TCN) is used for processing to obtain the second abrasive wear component. With the second adhesive wear component ;

[0048] D3. Compare the separation accuracy and execution time of the two blind source separation algorithms in D1 and D2; prioritize the abrasive wear component and adhesive wear component output by the blind source separation algorithm with higher separation accuracy as the final wear component; if the separation accuracy of the two blind source separation algorithms is the same, select the abrasive wear component and adhesive wear component output by the blind source separation algorithm with shorter execution time as the final wear component.

[0049] The separation accuracy is characterized by the kurtosis difference between the abrasive wear component and the adhesive wear component:

[0050] ;

[0051] ;

[0052] in, The kurtosis value of the wear component to be analyzed. The wear component to be analyzed is... Let be the mean of all wear components to be analyzed. To utilize the separation accuracy characterized by kurtosis, This represents the kurtosis value of the abrasive wear component. This represents the kurtosis value of the adhesive wear component.

[0053] Furthermore, data-level fusion is performed using the restored passive shear wave acoustic emission signal and the preprocessed active shear wave echo signal as joint inputs, including: using db4 wavelet decomposition to restore the passive shear wave acoustic emission signal. and preprocessed active shear wave echo signal Perform four-level wavelet decomposition and reconstruct the time-frequency energy spectrum. Then, concatenate all time-frequency features along the original signal dimension to obtain a data-level fused feature vector. ;

[0054] Feature-level fusion is performed using the separated wear components as input, including: merging the abrasive wear components filtered in step D3. With adhesive wear component Stack along the channel dimension to construct a two-channel input tensor Inputting into a one-dimensional convolutional neural network yields a high-dimensional abstract feature vector after feature-level fusion. The one-dimensional convolutional neural network includes: a first convolutional layer with a kernel size of 3 and a stride of 1; a second convolutional layer with a kernel size of 5 and a stride of 1; a max pooling layer with a window size of 2 and a fully connected layer; the first and second convolutional layers are both followed by the ReLU activation function, and a Dropout layer is connected after the pooling layer.

[0055] Furthermore, a wear depth regression network is used to fuse feature vectors at the data level. Feature vector fusion with feature level By concatenating the features, a multi-dimensional fused feature vector is obtained. As input, a Gaussian process regression model is used to establish a nonlinear mapping relationship between the fused feature vector and the wear depth, including:

[0056] E1. The training sample set is obtained offline. ,in The training data consists of the fused feature matrix formed by the data-level fusion result and the feature-level fusion result output in step S5 during the offline process. Let be the d-dimensional fusion feature vector of the i-th sample. The corresponding wear depth benchmark value, where N is the total number of training samples; wear depth benchmark value Obtained through surface three-dimensional morphology measurement;

[0057] E2, using multi-dimensional feature vector fusion For new test samples Based on the training samples and the test samples, calculate the covariance matrix K between the training samples and the covariance matrix between the test samples and each training sample. and the autocovariance matrix of the test samples The covariance matrix K is The square matrix, the element in the i-th row and j-th column , Let represent the radial basis function kernel function, and let represent the covariance between the i-th and j-th training samples; the covariance matrix represents the variance between the i-th and j-th training samples. For N e The vector of size ×1, the i-th element , representing the covariance between the test sample and the i-th training sample; the autocovariance matrix of the test sample. As a scalar,

[0058] E3. Based on the posterior distribution formula of a Gaussian process, calculate the predicted mean and predicted variance of the test sample:

[0059] Given training samples For test samples The posterior distribution of its predicted output is:

[0060] ;

[0061] Then predict the mean and prediction variance They are respectively:

[0062] ;

[0063] ;

[0064] in, It is the identity matrix. , This represents the noise variance.

[0065] Furthermore, the wear type classifier employs a multilayer perceptron to fuse feature vectors across multiple dimensions. Using the input as input, the classification prediction probability distribution is calculated through forward propagation. , This represents the predicted probability of abrasive wear. This represents the predicted probability of adhesive wear.

[0066] Step S6 further includes: predicting the mean As an average estimate of the wear depth under current operating conditions The average estimated value of wear depth Mapping to depth levels according to a preset five-level threshold. These correspond to normal, mild, moderate, severe, and extremely severe, respectively; the predicted variance is... As a morphological characteristic index of wear, the morphological characteristic index is used... Mapping amplitude magnitude to topography level The category corresponding to the highest predicted classification probability is used as the mechanism level. The depth level, morphology level, and mechanism level are weighted and summed according to preset weights to map to the final comprehensive wear level, including normal, slight, moderate, severe, and extremely severe.

[0067] The confidence level of the final comprehensive wear rating is calculated and determined jointly by the regression prediction variance and the classification prediction probability.

[0068] ;

[0069] in, To determine the confidence level of the final overall wear rating, These are the weight coefficients of the regression term. For the weight coefficients of the classification items, , To predict the upper limit of variance.

[0070] A transverse wave active-passive coordinated sensor includes a sleeve, which is a hollow metal cylindrical shell. An active element and a passive element are arranged at one end of the sleeve. The passive element is a broadband shear mode piezoelectric ring with a central circular hole in the middle. The active element is a high-frequency shear mode piezoelectric disk, which is arranged in the central through hole of the piezoelectric ring. A sound-absorbing backing is arranged at the other end of the sleeve and is positioned opposite to the sound wave emission direction of the piezoelectric element composed of the piezoelectric ring and the piezoelectric disk.

[0071] The beneficial effects of this invention are as follows:

[0072] (1) This invention uses the active shear wave echo signal as the reference input and employs an adaptive filtering algorithm to estimate and eliminate the modulation interference components in the passive shear wave acoustic emission signal. This effectively eliminates the vibration modulation effect of the active ultrasonic high-frequency disturbance on the passive AE signal, and obtains a pure passive shear wave acoustic emission restoration signal. This method solves the technical problem of passive AE signal distortion caused by the introduction of additional vibration components by active ultrasonic excitation, and significantly optimizes the accuracy and reliability of wear mechanism discrimination.

[0073] (2) This invention performs joint blind source separation processing on the restored passive AE signal and active shear wave echo signal, and adopts a feature extraction-based blind source separation method or a neural network-based end-to-end separation method to effectively decouple and independently output the signal components corresponding to abrasive wear and adhesive wear under coexisting conditions. This technology breaks through the bottleneck of signal source aliasing under composite wear mechanisms, which makes it difficult to distinguish multiple wear mechanisms, and realizes accurate separation and independent characterization of wear mechanism features.

[0074] (3) This invention proposes a strategy combining data-level fusion and feature-level fusion to efficiently fuse wear depth information acquired by active shear wave and wear mechanism information acquired by passive AE signal. Simultaneously, a wear depth regression algorithm is introduced to dynamically correct the active measurement depth based on the surface three-dimensional morphology measurement data. This method overcomes the limitations of traditional single longitudinal wave measurement technology, which can only acquire average thickness information and is difficult to perceive local morphology changes and wear mechanism transformations. Thus, it achieves coordinated, high-precision, real-time online characterization of the three dimensions of wear depth, surface morphology, and wear mechanism. Attached Figure Description

[0075] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0076] Figure 1 This is a system block diagram of the present invention for co-monitoring material wear based on transverse wave active ultrasound and passive acoustic emission.

[0077] Figure 2 This is an overall flowchart of the method for co-monitoring material wear based on transverse wave active ultrasound and passive acoustic emission according to the present invention.

[0078] Figure 3 This is a flowchart of the active and passive signal synchronous acquisition, signal restoration and blind source separation of the present invention.

[0079] Figure 4 This is a flowchart of the data fusion and wear characterization process of the present invention.

[0080] Figure 5 This is a schematic diagram of the active-passive cooperative sensor of the present invention.

[0081] Among them, 1 is the sleeve, 2 is the sound-absorbing backing, 3 is the passive element, and 4 is the active element. Detailed Implementation

[0082] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0083] Currently, most existing ultrasonic wear monitoring methods employ piezoelectric sensors in longitudinal (compression) wave mode. However, the propagation characteristics of longitudinal waves in solid-liquid-solid multilayer media mean that their response within the effective coverage area of ​​the sensor is primarily a comprehensive change in the average thickness of the oil film or coating. This single average thickness information has significant technical limitations in practical applications: longitudinal wave detection results reflect the average thickness within the coverage area, easily masking severe wear characteristics in localized areas, leading to measurement lag and difficulty in timely capturing the occurrence and evolution of early wear; under mixed lubrication or elastohydrodynamic lubrication conditions, early wear mainly manifests as changes in local geometry such as micro-scratches and micro-pits, and longitudinal waves lack sufficient sensitivity to such subtle geometric changes, making it difficult to effectively distinguish different wear stages or identify changes in wear mechanisms. To address these issues, piezoelectric sensors in transverse (shear) wave mode have significant advantages in wear monitoring. Transverse waves are more sensitive to changes in interfacial shear stiffness and contact state, and can more effectively reflect the local damage characteristics and contact mechanical behavior of the friction pair surface, possessing natural technical adaptability in early wear identification and local anomaly localization. Therefore, constructing a monitoring system based on the coordinated operation of active shear wave ultrasonic detection and passive shear wave acoustic emission, and solving problems such as vibration modulation interference of active shear wave on passive AE signal, signal decoupling and separation under composite wear mechanism, and fusion and regression correction of multi-source information on wear depth, is of great significance for improving the sensitivity and reliability of online wear monitoring.

[0084] A method for co-monitoring material wear based on transverse wave-based active ultrasound and passive acoustic emission, with the system architecture as follows: Figure 1 As shown, the method flow is as follows: Figure 2 As shown, it includes:

[0085] S1: Using the same clock reference, the active shear wave echo signal and the passive shear wave acoustic emission signal are synchronously acquired through a combined active and passive shear wave sensor.

[0086] The transverse wave active-passive coordinated sensor used in this invention, such as... Figure 5As shown, it includes a sleeve 1, which is a hollow metal cylindrical shell that serves to encapsulate, position, and provide electromagnetic shielding. One end of the sleeve 1 is equipped with an active element 4 and a passive element 3. The passive element 3 is a broadband shear mode piezoelectric ring with a central circular hole in the middle. The active element 4 is a high-frequency shear mode piezoelectric disk, which is arranged in the central through hole of the piezoelectric ring. The other end of the sleeve 1 is equipped with a sound-absorbing backing 2, which is located on the back of the piezoelectric element composed of the piezoelectric ring and the piezoelectric disk in the direction of sound wave emission.

[0087] In this application embodiment, a transverse wave active-passive cooperative sensor is proposed, such as... Figure 5 As shown, the sensor includes a sleeve 1, which is a hollow metal cylindrical shell that serves for encapsulation, positioning, and electromagnetic shielding. An active element 4 and a passive element 3 are concentrically arranged at one end of the sleeve 1. The passive element 3 is a broadband shear-mode piezoelectric ring with a frequency range of 1MHz to 10MHz. The passive element 3 primarily functions as a receiver, utilizing the positive piezoelectric effect to convert the mechanical vibration of the wear transverse wave acoustic emission from the workpiece into an electrical signal output. The active element 4 is a high-frequency shear-mode piezoelectric disk with a center frequency of 5MHz, arranged within the central through-hole of the piezoelectric ring. It converts electrical excitation into shear vibration through the inverse piezoelectric effect, emitting transverse waves towards the workpiece. It receives the transverse wave echo reflected from the workpiece interface through the positive piezoelectric effect. The active element 4 and the passive element 3 have their acoustic wave emission directions opposite to each other, and a sound-absorbing backing 2 is located at the other end of the sleeve 1. The vibration of the piezoelectric crystal will simultaneously emit transverse waves towards the workpiece (forward) and towards the inside of the sensor (backward); the backing absorbs and dissipates all the ultrasonic energy propagating backward, eliminating the trailing ringing noise caused by reflection from the back of the crystal, widening the sensor bandwidth, and avoiding noise contamination of active echo and passive acoustic emission signals.

[0088] When using, Figure 2 The active-passive coordinated sensor shown is fixedly mounted on the non-wear surface of the workpiece under test. It synchronously acquires the ultrasonic transverse wave signal emitted by the active element 4 and the transverse wave acoustic emission signal received by the passive element 3 during the wear process, using a sampling rate of 50MHz and the same clock reference, ensuring strict alignment of the two signals on the time axis. Subsequently, a time-domain synchronized active transverse wave echo signal sequence is output. (n is the sampling point index) and the passive transverse wave acoustic emission signal sequence This is used for subsequent preprocessing and modulation analysis.

[0089] S2: Preprocess the active shear wave echo signal and the passive shear wave acoustic emission signal, and analyze the modulation effect of the active shear wave echo signal on the passive shear wave acoustic emission signal, such as... Figure 2 As shown.

[0090] In this embodiment of the application, the preprocessing of the active shear wave echo signal and the passive shear wave acoustic emission signal includes: preprocessing the acquired active shear wave echo signal... With passive transverse wave acoustic emission signal Time-domain alignment and amplitude normalization preprocessing are performed to obtain the preprocessed active shear wave echo signal. and preprocessed passive transverse wave acoustic emission signal .

[0091] In this embodiment of the application, the analysis of the modulation effect of the active shear wave echo signal on the passive shear wave acoustic emission signal includes:

[0092] Based on the preprocessed active shear wave echo signal The amplitude attenuation characteristics and flight time variations are analyzed to extract parameters characterizing the interface contact stiffness changes within the sensor coverage area. The reference amplitude under no-wear conditions is assumed to be... The current echo amplitude is Interface stiffness variation parameters Defined as:

[0093] ;

[0094] Based on Hertzian contact theory, a mapping function F is established between the interface stiffness change and the modulation factor, mapping Δk[n] to the modulation factor:

[0095] ;

[0096] right Spline interpolation is performed to adjust the sampling rate of the preprocessed passive transverse wave acoustic emission signal. Consistent, resulting in a continuous modulation factor sequence. This modulation factor sequence will serve as the dynamic weighting coefficients for constructing the reference input signal in subsequent path one adaptive filtering.

[0097] S3: The passive shear wave acoustic emission signal is restored using both adaptive filtering and neural network regression algorithms, and the optimal restored passive shear wave acoustic emission signal is selected, such as... Figure 3 As shown.

[0098] In this embodiment, two shear-mode piezoelectric sensors are used. When active pulse echo (PE) and passive acoustic emission (AE) operate simultaneously, the active signal is mixed into the passive AE channel. Therefore, the passive AE signal needs to be restored. This invention employs a dual-path approach to restore the passive shear wave acoustic emission signal. Specifically:

[0099] Path 1: Signal restoration based on adaptive filtering: Adaptive transverse filtering (FIR) and the LMS algorithm are used to restore the passive transverse wave acoustic emission signal. The specific method is as follows:

[0100] First, the preprocessed active shear wave echo signal Dynamic modulation is performed using a modulation factor c[n] to obtain the dynamically modulated active shear wave echo signal:

[0101] ;

[0102] Furthermore, the dynamically modulated active shear wave echo signal As a reference input, the passive transverse wave acoustic emission signal As the desired response, adaptive transverse filtering (FIR) is performed for each sampling point n of the passive transverse wave acoustic emission signal:

[0103] Calculate the filter output, i.e., the estimated modulation interference components:

[0104] ;

[0105] in, ; The reference input vector is of order L; This represents the convolution operation.

[0106] Calculate the error signal, i.e., the passive transverse wave acoustic emission signal reconstructed from the current sampling point:

[0107] ;

[0108] Furthermore, update the weight vector according to the least mean square criterion:

[0109] ;

[0110] ;

[0111] in, The step size factor (using normalized LMS) is used, and the convergence coefficient α is set to 0.05~0.3 to compromise between convergence speed and steady-state error; the regularization minimum constant δ is a small constant used to prevent division by zero.

[0112] Furthermore, the error signal after the final iteration is taken as the first passive transverse wave acoustic emission signal for signal reconstruction based on adaptive filtering. .

[0113] Path Two: Signal Restoration Based on Neural Network Regression: A temporal convolutional network (TCN) with end-to-end regression is used as the signal restoration model to restore the passive shear wave acoustic emission signal. The specific method is as follows:

[0114] First, the dynamically modulated active shear wave echo signal and preprocessed passive transverse wave acoustic emission signal The synchronous timing segment constructs the input dual-channel timing signal The input window length is N, which is 1024 points (corresponding to a time window of approximately 20.48μs at a sampling rate of 50MHz in the embodiment). The input dimension of each channel is N, and the total input dimension is 2×N.

[0115] Furthermore, a temporal convolutional network (TCN) is used to process the dual-channel time-series signal. The process is performed to output the estimated modulation interference components. Temporal convolutional networks (TCNs) employ an encoder-decoder structure, specifically including the following processing steps:

[0116] Input layer: Receives dual-channel timing signals The window length is N=1024, and the output is a tensor with a dimension of 2×1024.

[0117] The first convolutional layer uses a one-dimensional convolutional kernel with a kernel size of 7, a stride of 1, and 32 output channels. The activation function is a rectified linear unit (ReLU). The feature map output by this layer has a dimension of 32×1024.

[0118] The residual block stack layer consists of six residual blocks connected in series. Each residual block contains a dilated convolution layer, a weight normalization layer, a ReLU activation function, and skip connections. The dilation rates of the dilated convolutions in each residual block are 1, 2, 4, 8, 16, and 32, respectively. The kernel size is 3, and the number of output channels remains at 32. The structure of each residual block is as follows:

[0119] ;

[0120] in, For dilated convolution, h in h is the input for the current residual block. out This is the output of the current residual block. The input of the first residual block is the output of the first convolutional layer. Subsequent residual blocks use the output of the current block as the input of the next block. The last residual block outputs a feature map with a dimension of 32×1024.

[0121] Output convolutional layer: Employs a one-dimensional convolutional kernel with a kernel size of 3, a stride of 1, and 1 output channel, with no activation function. This layer maps the 32-channel feature map to a single-channel temporal signal, outputting the estimated modulation interference component. The dimension is 1×1024.

[0122] Set up a subtraction layer: This will reduce the passive transverse wave acoustic emission signal. With the estimated modulation interference components By subtracting point by point, the second passive transverse wave acoustic emission signal based on neural network regression is obtained. .

[0123] During the network training phase, the mean square error between the restored signal and the target clean signal is used as the loss function, and its expression is:

[0124] ;

[0125] Where N is the window length of a single training sample, The training set contains known pure passive acoustic emission signals (acquired individually under conditions of no active emission).

[0126] The network was trained using the Adam optimizer, with an initial learning rate of 1×10⁻⁶. −3 The batch size is 32, and the training rounds are 200. After training, the preprocessed active shear wave echo signal is... and preprocessed passive transverse wave acoustic emission signal The synchronous time sequence segment is input into the network, and the second passive transverse wave acoustic emission signal is obtained by the network forward propagation calculation based on neural network regression. .

[0127] In this embodiment of the application, the optimal restored passive transverse wave acoustic emission signal is selected, including:

[0128] Different recovery methods are used for different operating conditions, and calibration is required first, that is, using active shear wave echo signals. As a reference signal, the passive transverse wave acoustic emission signal after the two paths are restored ( ) and reference signal To compare the two, calculate the amplitude squared coherence coefficient:

[0129] ;

[0130] In the formula, The cross-power spectral density between the restored passive signal and the active shear wave signal. and These are the self-power spectral densities of the restored passive signal and the active shear wave signal, respectively.

[0131] The average coherence coefficient is calculated within the effective frequency band of the passive shear wave acoustic emission sensor (1MHz to 10MHz in this embodiment):

[0132] ;

[0133] In the formula, the effective frequency band lower limit upper limit of effective frequency band .

[0134] By comparing the average coherence coefficients of the two paths, a smaller coherence coefficient indicates fewer residual active modulation interference components in the restored signal, resulting in a better restoration effect. The restoration result corresponding to the path with the smaller average coherence coefficient is selected as the final restored passive shear wave acoustic emission signal. :

[0135] ;

[0136] Will The signal is output to step S4 as the input signal for blind source separation.

[0137] S4: To restore the passive transverse wave acoustic emission signal and preprocessed active shear wave echo signal For joint input, the separated wear components are obtained through the following two blind source separation paths, such as... Figure 3 As shown.

[0138] In this embodiment of the application, the restored passive transverse wave acoustic emission signal output in step S3 is used. and active shear wave echo signal For joint input The FastICA algorithm is used for joint blind source separation to decouple the signal components corresponding to abrasive wear and adhesive wear, thus obtaining the abrasive wear component. With adhesive wear component .

[0139] Alternatively, an end-to-end blind source separation approach based on a Temporal Convolutional Network (TCN) can be used. This network is designed to reconstruct passive transverse wave acoustic emission signals. and preprocessed active shear wave echo signal As a dual-channel input, after processing by the causal convolution and dilated convolution modules of the TCN, the abrasive wear component is directly output. With adhesive wear component This method eliminates the need for explicit extraction of time-frequency domain statistical features, automatically learning the end-to-end mapping from aliased signals to each wear component through a network.

[0140] Compare the separation accuracy and execution time of the two blind source separation algorithms based on the wear component data obtained from the above steps:

[0141] (1) Separation accuracy: Calculate the output components of the two paths respectively. , , and The kurtosis for any signal component The formula for calculating its kurtosis is:

[0142]

[0143] In the formula, The signal component to be analyzed, Let M be the mean of the signal to be analyzed, and M be the signal length. The acoustic emission signal generated by abrasive wear is characterized by sparse, burst-type pulses with a high kurtosis value; while the acoustic emission signal generated by adhesive wear is characterized by a relatively continuous friction signal with a low kurtosis value.

[0144] The kurtosis difference between the abrasive wear component and the adhesive wear component for both paths was calculated separately to quantify the separation effect.

[0145]

[0146]

[0147] In the formula, This represents the kurtosis value of the abrasive wear component. This represents the kurtosis value of the adhesive wear component. A larger kurtosis difference indicates a more distinct distinction between the signal characteristics of the two wear mechanisms, resulting in better separation and higher accuracy.

[0148] (2) Execution time: Record the calculation time T for path one (FastICA) respectively. ICA The computation time T for path 2 (TCN) TCN To evaluate the real-time performance of the two paths.

[0149] Compare and Size:

[0150]

[0151] like = (If the separation accuracy of the two paths is comparable), then the path with the shorter execution time is selected:

[0152]

[0153] Based on the optimal selection result, output the abrasive wear component of the selected path. With adhesive wear component This serves as the input for data-level fusion and feature-level fusion in step S5.

[0154] S5: To restore the passive transverse wave acoustic emission signal and preprocessed active shear wave echo signal To perform data-level fusion of the combined inputs, the separated abrasive wear components output in step S4 are used. With adhesive wear component Perform feature-level fusion on the input, such as Figure 4 As shown,

[0155] In this embodiment of the application, the passive transverse wave acoustic emission signal is restored. and preprocessed active shear wave echo signal Perform data-level fusion for joint inputs, including:

[0156] The passive transverse wave acoustic emission signal was recovered using db4 wavelet decomposition. and active shear wave echo signal Perform four-level wavelet decomposition and reconstruct the time-frequency energy spectrum. Then, concatenate all time-frequency features along the original signal dimension to obtain a data-level fused feature vector. ;

[0157] In this embodiment of the application, the separated abrasive wear component output in step S4 is used. With adhesive wear component As a dual-channel input, the input undergoes feature-level fusion, including:

[0158] First, the abrasive wear component With adhesive wear component Stack along the channel dimension to construct a two-channel input tensor This is used as input to a one-dimensional convolutional neural network (1D CNN). This stacking method aligns the two wear components on the same time scale, making it easier for the network to learn the temporal correlation and independent features between the two components simultaneously.

[0159] The one-dimensional convolutional neural network, used as a feature fusion network, comprises the following structure: a first convolutional layer (kernel size 3, stride 1, output channels 16), a second convolutional layer (kernel size 5, stride 1, output channels 32), a max-pooling layer (window size 2), and a fully connected layer (output dimension 128). Both convolutional layers use the ReLU activation function, and a Dropout layer (dropout rate set to 0.5) is applied after the pooling layer to prevent overfitting.

[0160] During network forward propagation, dual-channel input The algorithm sequentially passes through a first convolutional layer, a second convolutional layer, and a max-pooling layer to extract multi-scale local temporal features. Then, a fully connected layer maps the pooled features into a 128-dimensional feature vector, which is activated by ReLU before being output. This output directly serves as the high-dimensional abstract feature vector after feature-level fusion. .

[0161] Data-level fusion of feature vectors Feature vector fusion with feature level By combining these features, we obtain the final multi-dimensional fused feature vector. This serves as the input for step S6.

[0162] S6: Fuse feature vectors at the data level Feature vector fusion with feature level Collaborative representation yields a multi-dimensional fused feature vector. As input, the wear depth regression network outputs an estimated wear depth and surface morphology feature index under the current operating conditions, and the wear mechanism category is output by the wear type classifier, such as... Figure 4 As shown.

[0163] In this embodiment, the wear depth regression network fuses feature vectors at the data level. Feature vector fusion with feature level By concatenating the features, a multi-dimensional fused feature vector is obtained. As input, a Gaussian Process Regression (GPR) model is used to establish a nonlinear mapping relationship between the fused feature vector and the wear depth.

[0164] Specifically, the method is as follows:

[0165] First, let the training sample set be... ,in The training data consists of the fused feature matrix formed by the data-level fusion result and the feature-level fusion result output in step S5 during the offline process. Let be the d-dimensional fusion feature vector of the i-th sample. is the corresponding wear depth benchmark value, and N is the total number of training samples.

[0166] The wear depth reference value The initial wear depth was obtained from surface 3D topography measurement data; the initial wear depth was obtained from preprocessed active shear wave echo signals. The average depth of the wear zone is determined based on measurements of flight time variations, implicitly contained in a multi-dimensional fused feature vector. In the subsequent process, the initial measured depth is corrected using a Gaussian process regression model before being output.

[0167] Furthermore, define the GPR model and select the kernel function:

[0168] Gaussian process regression model assumes target value With the input feature F (the multi-dimensional fused feature vector output from step S5) The following conditions must be met between )

[0169] ;

[0170] In the formula, f(F) is an unknown latent function, and ε is a function that follows a function with a mean of 0 and a variance of ε. The independent noise terms of the Gaussian distribution, This represents the noise variance.

[0171] Assign a Gaussian process prior to the latent function f(F):

[0172] ;

[0173] In the formula, m(F) is the mean function (in this embodiment, m(F) = 0). is the covariance function (i.e., the kernel function), used to measure the similarity between any two input samples F and F′.

[0174] In this embodiment, the kernel function used is the Radial Basis Function (RBF), which is defined as follows:

[0175] ;

[0176] In the formula, is the signal variance, controlling the magnitude of the function; l is the length scale parameter, controlling the smoothness of the function's changes; ||FF′|| is the Euclidean distance between the two input vectors.

[0177] Furthermore, based on the training samples and the test samples (the fusion features obtained online), the covariance matrix K between the training samples and the covariance matrix between the test samples and each training sample are calculated. and the autocovariance matrix of the test samples The covariance matrix K is A square matrix, where N is the number of training samples, and the element in the i-th row and j-th column is... Let represent the covariance between the i-th and j-th training samples, and let be the covariance matrix. For N e The vector of size ×1, the i-th element , representing the covariance between the test sample and the i-th training sample, the covariance matrix. As a scalar, .

[0178] Furthermore, based on the posterior distribution formula of the Gaussian process, the predicted mean and predicted variance of the test sample are calculated.

[0179] Given training samples For new test samples (Fusion features obtained online) The posterior distribution of its predicted output is:

[0180] ;

[0181] Forecast Mean and prediction variance They are respectively:

[0182] ;

[0183] ;

[0184] In the formula, I represents the predicted output corresponding to the test sample, i.e., the corrected predicted wear depth value output by the model, where I is N. e ×N e The identity matrix, This represents the noise variance. The predicted mean is... This is the average estimated wear depth under the current operating conditions output by the model. Prediction variance Provide a measure of the uncertainty (confidence interval) of the estimate.

[0185] In this embodiment, the wear type classifier uses a multilayer perceptron (MLP) to jointly define feature vectors. ∈R d The network structure, taking input as input, includes an input layer (dimension d), two hidden layers (64 and 32 neurons respectively, both with ReLU activation function), and an output layer (2 neurons, Softmax activation). The classifier calculates the predicted probability through forward propagation.

[0186]

[0187]

[0188]

[0189]

[0190] In the formula, W1, W2, and W3 are the weight matrices of each layer, and b1, b2, and b3 are the bias vectors of each layer. The predicted probability distributions for abrasive wear and adhesive wear satisfy the following conditions: , where e is the base of the natural logarithm.

[0191] After processing, the mean was predicted using Gaussian process regression. As the corrected average estimate of wear depth To predict variance As a morphological characteristic index, the corrected average estimate of wear depth is used. The classification prediction probability p (including wear mechanism category and confidence level) and morphological feature index are used as three dimensions to construct a multi-dimensional quantitative assessment model of overall wear status.

[0192] The overall wear level is obtained by weighted fusion of three dimensions: first, the wear depth value... Mapping to depth levels according to a preset five-level threshold. (Corresponding to normal, mild, moderate, severe, and extremely severe, respectively); morphological characteristic index Mapping amplitude magnitude to topography level The category corresponding to the highest classification probability is taken as the mechanism level L. c (Abrasive wear or adhesive wear). Then, the three levels are weighted and summed according to preset weights to map to the final comprehensive wear level (five levels in total: normal, slight, moderate, severe, and extremely severe).

[0193] The confidence interval is determined jointly by the regression prediction variance and the classification prediction probability:

[0194] ;

[0195] In the formula, , This is the upper limit of variance determined during the calibration phase. This confidence level reflects the reliability of the current evaluation results; the higher the confidence level, the more reliable the regression prediction and classification judgment.

[0196] Finally, the comprehensive wear level and confidence interval under the current working condition are output in the form of a three-dimensional vector [wear depth, morphological feature index, wear mechanism category], realizing quantitative and multi-angle online evaluation of wear status.

[0197] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for co-monitoring material wear based on transverse wave-based active ultrasound and passive acoustic emission, characterized in that, include: S1: Using the same clock reference, the active shear wave echo signal and the passive shear wave acoustic emission signal are synchronously acquired through the active and passive shear wave co-sensor. S2: Preprocess the active shear wave echo signal and the passive shear wave acoustic emission signal, and analyze the modulation effect of the active shear wave echo signal on the passive shear wave acoustic emission signal. S3: The passive transverse wave acoustic emission signal is restored by using adaptive filtering and neural network regression algorithms respectively, and the optimal restored passive transverse wave acoustic emission signal is selected. S4: Using the restored passive transverse wave acoustic emission signal and the preprocessed active transverse wave echo signal as joint inputs, the separated wear component is obtained through blind source separation. S5: Data-level fusion is performed using the restored passive transverse wave acoustic emission signal and the preprocessed active transverse wave echo signal as joint inputs, and feature-level fusion is performed using the separated wear component as input. S6: Using the multi-dimensional fused feature vector obtained by the collaborative representation of data-level fused feature vector and feature-level fused feature vector as input, the wear depth regression network outputs the wear depth estimate and surface morphology feature index under the current working condition, and the wear mechanism category is output through the wear type classifier.

2. The method for co-monitoring material wear based on transverse wave-driven active ultrasound and passive acoustic emission according to claim 1, characterized in that, Analysis of the modulation effect of active shear wave echo signal on passive shear wave acoustic emission signal, including: A1. Based on the preprocessed active shear wave echo signal The amplitude attenuation characteristics and flight time variations are used to extract parameters characterizing the interface contact stiffness changes within the sensor coverage area. ； in, The reference amplitude under no-wear condition is, This represents the echo amplitude of the preprocessed active shear wave echo signal. A2. Based on Hertzian contact theory, establish a mapping function F between interface stiffness variation and modulation factor, and characterize the interface contact stiffness variation as a parameter. Mapped to modulation factor: ； A3. Modulation factor Perform spline interpolation to align the preprocessed passive transverse wave acoustic emission signal. Sampling rate, to obtain a continuous modulation factor sequence .

3. The method for co-monitoring material wear based on transverse wave-driven active ultrasound and passive acoustic emission according to claim 2, characterized in that, Signal restoration based on adaptive filtering includes: restoring passive transverse wave acoustic emission signals using adaptive transverse filtering (FIR) and LMS algorithms. B1. The preprocessed active shear wave echo signal Dynamic modulation is performed using a continuous modulation factor c[n] to obtain the dynamically modulated active shear wave echo signal: ; B2. The dynamically modulated active shear wave echo signal As a reference input, the passive transverse wave acoustic emission signal As the desired response, for each sampling point n of the passive transverse wave acoustic emission signal, an adaptive transverse filtering FIR iterative calculation based on the LMS algorithm is performed: B21. Calculate the estimated modulation interference components: ； in, ; The reference input vector is of order L; This represents the convolution operation. B22. Calculate the error signal, i.e., the passive transverse wave acoustic emission signal reconstructed from the current sampling point: ； B23. Update the weight vector according to the LMS (Least Mean Square) criterion: ； ； in, δ is the step size factor, α is the convergence coefficient used to compromise between convergence speed and steady-state error; δ is the regularization minimum constant used to prevent division by zero. B3. Select the error signal after the final iteration as the first passive transverse wave acoustic emission signal for signal restoration based on adaptive filtering. .

4. The method for co-monitoring material wear based on transverse wave-driven active ultrasound and passive acoustic emission according to claim 3, characterized in that, Signal restoration based on neural network regression includes: using dynamically modulated active shear wave echo signals. and preprocessed passive transverse wave acoustic emission signal The synchronous timing segment constructs the input dual-channel timing signal Based on the first temporal convolutional network TCN, dual-channel time-series signals The signal was processed to obtain the second passive transverse wave acoustic emission signal based on neural network regression. The first temporal convolutional network (TCN) adds a subtraction layer to the original TCN network to process the preprocessed passive transverse wave acoustic emission signal. Modulation interference components estimated by the original TCN network By subtracting point by point, the second passive transverse wave acoustic emission signal based on neural network regression is obtained. .

5. The method for co-monitoring material wear based on transverse wave-driven active ultrasound and passive acoustic emission according to claim 4, characterized in that, Selecting the optimal restored passive transverse wave acoustic emission signal includes: C1. Using the pre-processed active shear wave echo signal The passive transverse wave acoustic emission signal after the two paths are restored, serving as a reference signal. Respectively with reference signal To compare the two paths, the amplitude squared coherence coefficients between the restored passive transverse wave acoustic emission signals and the reference signal were calculated: ； In the formula, The cross-power spectral density between the restored passive signal and the active shear wave signal. and These are the autopower spectral densities of the restored passive signal and the reference signal, respectively. C2. Calculate the average coherence coefficients of the passive shear wave acoustic emission signals and the reference signal after restoration for both paths within the effective frequency band of the passive shear wave acoustic emission sensor: ； In the formula, As the lower limit of the effective frequency band, f H This is the upper limit of the effective frequency band; C3. Compare the average coherence coefficients of the two restoration paths, and select the restoration signal corresponding to the one with the smaller average coherence coefficient as the final restored passive transverse wave acoustic emission signal. : 。 6. The method for co-monitoring material wear based on transverse wave-driven active ultrasound and passive acoustic emission according to claim 5, characterized in that, The wear component after separation is obtained through blind source separation, including: D1. To restore the passive transverse wave acoustic emission signal and preprocessed active shear wave echo signal For joint input The FastICA algorithm is used for joint blind source separation to decouple the signal components corresponding to abrasive wear and adhesive wear, thus obtaining the first abrasive wear component. With the first adhesive wear component ; D2. To restore the passive transverse wave acoustic emission signal and preprocessed active shear wave echo signal For joint input, a second temporal convolutional network (TCN) is used for processing to obtain the second abrasive wear component. With the second adhesive wear component ; D3. Compare the separation accuracy and execution time of the two blind source separation algorithms in D1 and D2; prioritize the abrasive wear component and adhesive wear component output by the blind source separation algorithm with higher separation accuracy as the final wear component; if the separation accuracy of the two blind source separation algorithms is the same, select the abrasive wear component and adhesive wear component output by the blind source separation algorithm with shorter execution time as the final wear component. The separation accuracy is characterized by the kurtosis difference between the abrasive wear component and the adhesive wear component: ； ； in, The kurtosis value of the wear component to be analyzed. The wear component to be analyzed is... Let be the mean of all wear components to be analyzed. To utilize the separation accuracy characterized by kurtosis, This represents the kurtosis value of the abrasive wear component. This represents the kurtosis value of the adhesive wear component.

7. The method for co-monitoring material wear based on transverse wave-driven active ultrasound and passive acoustic emission according to claim 6, characterized in that, Data-level fusion is performed using the restored passive shear wave acoustic emission signal and the preprocessed active shear wave echo signal as joint inputs, including: using db4 wavelet decomposition to restore the passive shear wave acoustic emission signal. and preprocessed active shear wave echo signal Perform four-level wavelet decomposition and reconstruct the time-frequency energy spectrum. Then, concatenate all time-frequency features along the original signal dimension to obtain a data-level fused feature vector. ; Feature-level fusion is performed using the separated wear components as input, including: merging the abrasive wear components filtered in step D3. With adhesive wear component Stack along the channel dimension to construct a two-channel input tensor Inputting into a one-dimensional convolutional neural network yields a high-dimensional abstract feature vector after feature-level fusion. The one-dimensional convolutional neural network includes: a first convolutional layer with a kernel size of 3 and a stride of 1; a second convolutional layer with a kernel size of 5 and a stride of 1; a max pooling layer with a window size of 2 and a fully connected layer; the first and second convolutional layers are both followed by the ReLU activation function, and a Dropout layer is connected after the pooling layer.

8. The method for co-monitoring material wear based on transverse wave-driven active ultrasound and passive acoustic emission according to claim 7, characterized in that, Wear depth regression network, fusing feature vectors at the data level Feature vector fusion with feature level By concatenating the features, a multi-dimensional fused feature vector is obtained. As input, a Gaussian process regression model is used to establish a nonlinear mapping relationship between the fused feature vector and the wear depth, including: E1. The training sample set is obtained offline. ,in The training data consists of the fused feature matrix formed by the data-level fusion result and the feature-level fusion result output in step S5 during the offline process. Let be the d-dimensional fusion feature vector of the i-th sample. The corresponding wear depth benchmark value, where N is the total number of training samples; wear depth benchmark value Obtained through surface three-dimensional morphology measurement; E2, using multi-dimensional feature vector fusion For new test samples Based on the training samples and the test samples, calculate the covariance matrix K between the training samples and the covariance matrix between the test samples and each training sample. and the autocovariance matrix of the test samples The covariance matrix K is The square matrix, the element in the i-th row and j-th column , Let represent the radial basis function kernel function, and let represent the covariance between the i-th and j-th training samples; the covariance matrix represents the variance between the i-th and j-th training samples. For N e The vector of size ×1, the i-th element , representing the covariance between the test sample and the i-th training sample; the autocovariance matrix of the test sample. As a scalar, E3. Based on the posterior distribution formula of a Gaussian process, calculate the predicted mean and predicted variance of the test sample: Given training samples For test samples The posterior distribution of its predicted output is: ； Then predict the mean and prediction variance They are respectively: ； ； in, It is the identity matrix. , This represents the noise variance.

9. The method for co-monitoring material wear based on transverse wave-driven active ultrasound and passive acoustic emission according to claim 7 or 8, characterized in that, The wear type classifier employs a multilayer perceptron to fuse feature vectors from multiple dimensions. Using the input as input, the classification prediction probability distribution is calculated through forward propagation. , This represents the predicted probability of abrasive wear. This represents the predicted probability of adhesive wear. Step S6 further includes: predicting the mean As an average estimate of the wear depth under current operating conditions The average estimated value of wear depth Mapping to depth levels according to a preset five-level threshold. These correspond to normal, mild, moderate, severe, and extremely severe, respectively; the predicted variance is... As a morphological characteristic index of wear, the morphological characteristic index is used... Mapping amplitude magnitude to topography level The category corresponding to the highest predicted classification probability is used as the mechanism level. The depth level, morphology level, and mechanism level are weighted and summed according to preset weights to map to the final comprehensive wear level, including normal, slight, moderate, severe, and extremely severe. The confidence level of the final comprehensive wear rating is calculated and determined jointly by the regression prediction variance and the classification prediction probability. ； in, To determine the confidence level of the final overall wear rating, These are the weight coefficients of the regression term. For the weight coefficients of the classification items, , To predict the upper limit of variance.

10. A transverse wave active-passive coordinated sensor, characterized in that, Includes a sleeve (1), which is a hollow metal cylindrical shell; an active element (4) and a passive element (3) are arranged at one end of the sleeve (1), the passive element (3) is a broadband shear mode piezoelectric ring with a central circular hole in the middle of the piezoelectric ring; the active element (4) is a high frequency shear mode piezoelectric disk, which is arranged in the central through hole of the piezoelectric ring; a sound-absorbing backing (2) is arranged at the other end of the sleeve (1), and is positioned opposite to the sound wave emission direction of the piezoelectric element composed of the piezoelectric ring and the piezoelectric disk.

Images