Engine alloy component wear test method and system based on acoustic emission
By independently clustering and generating composite cluster labels in the spectral time domain and energy domain feature spaces, and constructing a sample relationship graph by combining Euclidean distance and Riemann distance, the problem of insufficient accuracy in wear state identification in the prior art is solved, and high-precision wear state identification is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGZHOU TIANMA GRP TIANMA MOTORCYCLE CO LTD
- Filing Date
- 2026-04-28
- Publication Date
- 2026-05-29
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing acoustic wear detection methods suffer from insufficient accuracy in wear state identification due to feature fusion not taking into account multi-domain structural relationships and sample relationship measurement being limited to Euclidean distance. This is especially prone to misjudgment and missed judgment in the early stages of wear.
By extracting spectral time-domain and energy-domain features, clustering is performed independently in these two spaces to generate composite cluster labels. The local covariance matrix is calculated, and Euclidean distance and Riemann distance are combined and transformed into edge weights through a kernel function to construct a sample relationship graph. The label propagation algorithm is then used to identify the wear state under a small number of calibrated samples.
It achieves high-precision identification of wear state of engine alloy parts in scenarios where labeled samples are scarce, improves the accuracy and reliability of wear state identification, and eliminates the dominant effect of dimensional differences on subsequent distance measurement and covariance structure.
Smart Images
Figure CN122109335A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of acoustic signal processing and fault diagnosis technology. More specifically, this invention relates to a method and system for testing wear of engine alloy components based on acoustic emission. Background Technology
[0002] Engine alloy components operate under high load, high speed, and complex stress environments. Wear is one of the main causes of engine performance degradation and even sudden failures. Accurate monitoring and diagnosis of the wear condition of alloy components is of significant engineering importance for preventing failures, optimizing maintenance strategies, and extending equipment service life. Acoustic signal analysis, as a non-invasive condition monitoring technology, can remotely and online detect acoustic radiation generated by wear behaviors such as friction, impact, and crack propagation, providing a feasible technical approach for wear condition assessment.
[0003] The acoustic signals of an engine during actual operation are extremely complex, typically exhibiting non-stationary and nonlinear characteristics. Furthermore, the weak early wear characteristic signals are often masked by strong combustion noise, mechanical vibration noise, and environmental noise, resulting in an extremely low signal-to-noise ratio. This makes it difficult for traditional time-frequency analysis methods to reliably extract sensitive features closely related to the evolution of wear states. To address these difficulties, relevant technical solutions utilize nonlinear operators such as the Teager-Kaiser energy operator to enhance the transient impact components in the signal, highlighting abnormal vibrations caused by wear defects. Simultaneously, by extracting and fusing multi-domain features from the spectral time domain and energy domain, a more comprehensive representation of the signal state is achieved. Based on this, classification or clustering models are constructed to identify different wear levels.
[0004] However, the aforementioned technical approaches still face two limitations. In the feature fusion stage, a simple feature vector concatenation method is often used. This approach fails to consider the potential inherent structural relationships and complementary information between different feature domains, thus failing to achieve deep feature fusion and easily leading to information redundancy or dilution of key information. When constructing relationships between samples or performing cluster analysis, traditional metrics such as Euclidean distance are typically relied upon. These metrics have limitations in representing high-dimensional feature spaces, failing to accurately perceive the complex manifold structure inherent in the data, thereby affecting the accurate judgment of sample similarity. This lack of perception of the data's inherent geometric structure limits the model's ability to identify subtle differences in the evolution of wear states, especially in the early stages of wear, easily leading to misjudgments and missed judgments. Summary of the Invention
[0005] To address the technical problem that existing acoustic wear detection methods suffer from insufficient accuracy in wear state identification due to feature fusion not taking into account multi-domain structural correlations and sample relationship measurement being limited to Euclidean distance, this invention provides solutions in the following aspects.
[0006] In a first aspect, the present invention provides a method for testing the wear of engine alloy components based on acoustic emission, comprising: The system acquires continuous sound signals from the engine alloy components during operation and divides these signals into multiple data frames. The sound signals are acoustic emission signals acquired by acoustic emission sensors. For each data frame, a spectral time-domain feature set is extracted. The absolute value of the Teager-Kaiser energy operator is applied to the data frame to obtain an instantaneous energy sequence, and an energy domain feature set is extracted. Z-score normalization is performed on each dimension of the spectral time-domain feature set and the energy domain feature set. The normalized spectral time-domain feature set and the energy domain feature set are concatenated to form an initial combined feature set. All data frames are independently clustered in the spectral time-domain feature space and the energy-domain feature space, respectively. Spectral time-domain cluster labels and energy-domain cluster labels are generated for each data frame and paired into composite cluster labels. For each composite cluster label, the initial combined features of its subordinate data frames are collected to calculate the shared local covariance matrix. A sample relationship graph is constructed using data frames as nodes. For any two data frames, the Euclidean distance between their initial combined features and the Riemann distance between the local covariance matrices corresponding to the labels of their respective composite clusters are calculated. The Euclidean distance and the Riemann distance are input into a kernel function and fused to obtain the edge weights between them. The wear state of all data frames is determined by combining the calibration samples with known wear state labels on the sample relationship graph and using a label propagation algorithm. The wear test results of the engine alloy components are then output.
[0007] This invention generates composite cluster labels by independently clustering features in the spectral time domain feature space and energy domain feature space, respectively. Based on this, the local covariance matrix of each composite cluster is calculated. When constructing the sample relationship graph, the Euclidean distance between the initial combined features and the Riemann distance between the local covariance matrices are fused and transformed into edge weights through a kernel function. This allows the sample relationship graph to simultaneously capture the positional differences of data frames at the eigenvalue level and the geometrical differences at the covariance structure level. Combined with the label propagation algorithm, wear state information can be propagated along the graph structure to all unlabeled data frames under the condition of a small number of labeled samples. Thus, high-precision identification of wear state of engine alloy parts is achieved in scenarios where labeled samples are scarce.
[0008] Preferably, the step of dividing the acquired audio signal into multiple data frames includes: setting the frame length of the data frame to 2048 sampling points, the frame shift to 1024 sampling points to form a 50% overlap rate, and applying a Hamming window to each data frame to reduce spectral leakage.
[0009] Preferably, the step of extracting the spectral time-domain feature set for each data frame includes: For each data frame, the 13-dimensional Mel frequency cepstral coefficients, spectral centroid, spectral roll-off point, and zero-crossing rate are extracted, and these four features are combined to form a spectral time-domain feature set.
[0010] Extracting 13-dimensional Mel frequency cepstral coefficients, spectral centroid, spectral roll-off point, and zero-crossing rate to construct a spectral time-domain feature set can jointly characterize the acoustic state of the data frame from three aspects: spectral envelope shape, energy distribution centroid, high-frequency attenuation characteristics, and time-domain waveform noise content, thus improving the comprehensiveness of the spectral time-domain features in representing acoustic changes caused by wear.
[0011] Preferably, the step of applying the Teager-Kaiser energy operator to the data frame, taking the absolute value to obtain the instantaneous energy sequence, and extracting the energy domain feature set includes: The mean, standard deviation, and energy entropy of the instantaneous energy sequence within each data frame are calculated, and these three indicators are used as the energy domain feature set.
[0012] The mean, standard deviation, and energy entropy of the instantaneous energy sequence are calculated as the energy domain feature set, which enables the feature representation to capture the non-stationary transient impact components caused by wear failure, thus making up for the inadequacy of spectral time domain features in representing transient impact signals.
[0013] Preferably, the clustering algorithm is a density-based noise-based spatial clustering algorithm; during independent clustering, for each data point, the distance to its next cluster is calculated. The distances to the nearest neighbors are used to plot a k-distance graph, where all data points are sorted in ascending order of distance values. The distance value corresponding to the inflection point of the curve is selected as the neighborhood radius. Preset minimum sample size Based on the neighborhood radius and the minimum number of samples Perform clustering; the Values and Consistent.
[0014] The k-distance graph inflection point method is used to automatically determine the neighborhood radius and perform clustering based on this parameter, which avoids the subjectivity of manually preset clustering parameters and improves the adaptability of clustering results to different data distribution patterns.
[0015] Preferably, the calculation of the shared local covariance matrix includes: For those belonging to the same composite cluster label Initial combined feature vectors ( ), Calculate the mean vector: ; Local covariance matrix The relation is: ; In the formula, To belong to the first label under this composite cluster The initial combined feature vector of each data frame, This is the difference vector between the initial combined feature vector and the mean vector. Difference vector transpose, For regularization parameters, It is an identity matrix.
[0016] By calculating the shared local covariance matrix of the initial combined eigenvectors belonging to the same composite cluster label and introducing a regularization parameter to ensure the positive definiteness of the matrix, a numerically stable geometric basis is provided for the matrix inversion and matrix logarithm operations required in the subsequent Riemann distance calculation.
[0017] Preferably, the step of calculating the Euclidean distance between the initial combined features of any two data frames and the Riemann distance between the local covariance matrices corresponding to the composite cluster labels of the two data frames for any two data frames, and then fusing the Euclidean distance and the Riemann distance into a kernel function to obtain the edge weights between the two data frames, includes: For any two data frames, calculate the Euclidean distance between the initial combined features; For the local covariance matrix corresponding to the composite cluster labels of the two data frames and The Riemann distance is calculated using the affine-invariant Riemann metric. Edge weights are calculated using a Gaussian kernel function.
[0018] By combining Euclidean distance and affine-invariant Riemann distance and transforming them into edge weights using a Gaussian kernel function, the edge weights can simultaneously reflect the differences in eigenvalues between data frames and the differences in the covariance structure of their respective data clusters, thereby enhancing the accuracy of the sample relationship graph in representing the intrinsic differences of different wear states.
[0019] Preferably, determining the wear status of all data frames using the tag propagation algorithm includes: Construct the probability transition matrix probability transition matrix equal to a diagonal matrix The inverse matrix multiplied by the weight matrix Weight matrix The weight matrix and diagonal matrix of the sample relationship graph. The diagonal elements are the weight matrix. The sum of all elements in the corresponding row; Initialize the label matrix The row vectors corresponding to the calibrated samples are set as one-hot encoded vectors with known wear states, while the row vectors corresponding to the uncalibrated samples are initially set to all zeros. By iteratively updating the label matrix, the label matrix at the next time step is equal to the probability transition matrix. Multiply by the label matrix at the current time step, and reset the labels of the labeled samples to the known state after each iteration, until the label matrix is fully updated. Convergence determines the wear status of all data frames.
[0020] By iteratively updating the label matrix and resetting the labels of the labeled samples to the known state after each iteration, the data distribution information contained in a large number of unlabeled samples is fully utilized, and reliable determination of wear status is achieved under the condition of scarce labeled samples.
[0021] Preferably, the relationship of the Gaussian kernel function is: ; In the formula, For data frames and data frames Edge weights between them The Euclidean distance between the initial combined features. The Riemann distance between local covariance matrices. and The bandwidth parameter is a positive value.
[0022] Secondly, the present invention provides an engine alloy component wear testing system based on acoustic emission, comprising a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned engine alloy component wear testing method based on acoustic emission is implemented.
[0023] By adopting the above technical solution, the above-mentioned acoustic emission-based engine alloy component wear test method is generated into a computer program and stored in a memory so that it can be loaded and executed by a processor. A terminal device can then be made based on the memory and processor for convenient use.
[0024] The beneficial effects of this invention are as follows: This invention extracts spectral temporal features and energy domain features based on the Teager-Kaiser energy operator and concatenates them to form initial combined features. Independent clustering and pairing are then performed in both the spectral temporal and energy domain feature spaces to form composite cluster labels. Based on these composite cluster labels, the initial combined features of subordinate data frames are aggregated to calculate a shared local covariance matrix, thereby encoding the feature association structure within each data cluster. When constructing the sample relationship graph, the Euclidean distance and the Riemann distance between the local covariance matrices are fused and transformed into edge weights using a kernel function. This allows the graph model to simultaneously perceive positional differences at the feature value level and geometric differences at the covariance structure level. A label propagation algorithm then propagates wear state information along the graph structure to all data frames with a small number of calibrated samples. Dimensional Z-score standardization of the spectral temporal and energy domain features eliminates the dominant effect of dimensional differences on subsequent distance measurements and covariance structures. Furthermore, by comprehensively utilizing multi-domain feature information and the local geometric structure of the data, the accuracy of wear state identification is improved, making the test results more reliable.
[0025] Furthermore, the composite cluster label cross-encodes the clustering information of the two feature spaces, making data frames belonging to the same composite cluster label similar in spectral temporal structure and energy response mode. This achieves a more refined and physically meaningful local structural division of sound data, providing a reasonable grouping basis for the calculation of the local covariance matrix. Attached Figure Description
[0026] Figure 1 This is a flowchart illustrating the wear testing method for engine alloy components based on acoustic emission in this invention; Figure 2 This is a schematic diagram illustrating the cluster distribution of composite cluster labels; Figure 3 It is a bar chart that schematically shows the performance comparison of different schemes. Detailed Implementation
[0027] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.
[0028] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0029] This invention discloses a wear testing method for engine alloy components based on acoustic emission, referring to... Figure 1 This includes steps S1-S3: S1. Acquire continuous sound signals from the engine alloy components during operation, and divide the acquired sound signals into multiple data frames. The sound signals are acoustic emission signals acquired by acoustic emission sensors. Extract spectral time-domain feature sets from each data frame, apply the Teager-Kaiser energy operator to the data frame and take the absolute value to obtain the instantaneous energy sequence and extract the energy domain feature set. Perform Z-score normalization on each dimension of the spectral time-domain feature set and the energy domain feature set respectively, and cascade the normalized spectral time-domain feature set and the energy domain feature set to form an initial combined feature.
[0030] Acoustic emission signals from engine alloy components during operation are collected using a piezoelectric acoustic emission sensor with a frequency response range of 100kHz to 1MHz. The sampling rate is 1MHz, and the sampling bit depth is 16 bits. The collected data is stored in digital waveform format. The acquisition duration is determined according to actual diagnostic needs, and in this embodiment, it is no less than 3 minutes to ensure coverage of a sufficient number of operating cycles.
[0031] The acquired continuous audio signal is processed by frame segmentation. The frame length is set to 2048 sampling points, which corresponds to approximately 2.048 ms. The frame shift is 1024 sampling points, half the frame length, resulting in a 50% overlap rate. A Hamming window is applied to each data frame to reduce spectral leakage. The Hamming window is a well-known windowing function and will not be elaborated upon here.
[0032] For each segmented data frame, a spectral time-domain feature set is extracted. This set consists of four types of features: 13-dimensional Mel frequency cepstral coefficients, which are the first 13 coefficients obtained by taking the logarithmic energy of the power spectrum of each data frame after passing it through a Mel filter bank and performing a discrete cosine transform; the spectral centroid, which characterizes the centroid position of the spectral energy distribution; the spectral roll-off point, which is the frequency value corresponding to the accumulation of spectral energy reaching a preset percentage of the total energy; and the zero-crossing rate, which is the frequency at which the signal waveform crosses zero within the data frame. These four types of features are concatenated to form a 16-dimensional spectral time-domain feature vector.
[0033] It is worth noting that the Mel frequency cepstral coefficients can characterize the shape of the spectral envelope of the sound signal, the spectral centroid reflects the high and low frequency shift trend of the energy distribution, the spectral roll-off point characterizes the attenuation characteristics of high frequency energy, and the zero-crossing rate is sensitive to the noise content of the signal. Combining these four factors can jointly characterize the acoustic state of the data frame from three levels: spectral envelope, energy distribution, and time-domain waveform.
[0034] The instantaneous energy sequence is obtained by applying the Teager-Kaiser energy operator to the original audio signal within the data frame and taking the absolute value. The Teager-Kaiser energy operator is a nonlinear instantaneous energy estimation operator that enhances transient impulse components in a signal. For discrete signals... The discrete form of the Teager-Kaiser energy operator is: The instantaneous energy sequence is obtained by taking the absolute value of the operator output. .
[0035] Based on the instantaneous energy sequence, an energy domain feature set is extracted. This energy domain feature set consists of the following three statistical indicators: the mean of the instantaneous energy sequence. , representing the average instantaneous energy level within the data frame; the standard deviation of the instantaneous energy sequence. This indicates the intensity of instantaneous energy fluctuations; and energy entropy. .
[0036] The length is Data frames are divided into The non-overlapping subframes, the first Each subframe consists of an instantaneous energy sequence. The sampling points at the corresponding positions constitute the first... Energy of each subframe For all within this subframe The sum of sampled values; Normalize the energy of each subframe to obtain the probability: ; The relationship between energy and entropy is: ; According to the conventions of information theory ; In the formula, Energy entropy; In this embodiment, the number of subframes is... The value is 8, and in other embodiments it can be adjusted within the range of [4, 16] depending on the data frame length; For the first The energy of each subframe is determined by the instantaneous energy sequence. Belongs to the first It is calculated by summing the sampled values of each subframe; For the first The ratio of the energy of a subframe to the sum of the energy of all subframes.
[0037] Energy entropy This measures the uniformity of instantaneous energy distribution across subframes within a data frame. When energy is uniformly distributed across subframes, each... Approaching , Get the maximum value When energy is concentrated in a few subframes, the energy in the remaining subframes... Approaching zero Significantly reduced. In wear detection scenarios, the impact acoustic emission events generated by wear cause energy to suddenly concentrate in a specific period of time, reducing energy entropy, which can be used to distinguish between normal operation and wear conditions.
[0038] Before concatenation, Z-score standardization is performed on each dimension of the spectral time-domain feature vector and each dimension of the energy-domain feature vector. For the same dimension feature value across all data frames, the mean and standard deviation of that dimension are calculated. The original value of that dimension in each data frame is subtracted from the mean and then divided by the standard deviation to obtain the standardized feature value. The reason for performing standardization is that the dimensions of the spectral time domain and the dimensions of the energy domain differ significantly in terms of dimensions and orders of magnitude—the Mel frequency cepstral coefficients, spectral centroid, zero-crossing rate, and instantaneous energy mean, standard deviation, and energy entropy belong to different physical quantities. Direct concatenation would cause the high-order-order dimensions to dominate subsequent distance calculations and covariance structures, while the contribution of the low-order-order dimensions would be diluted. After Z-score standardization, the mean of each dimension is zero and the standard deviation is one, the dimensions are eliminated, and each dimension has equal weight in the initial combined feature vector. The standardized 16-dimensional spectral time-domain feature vector and the 3-dimensional energy-domain feature vector are concatenated sequentially to form a 19-dimensional initial combined feature vector, which serves as the complete feature representation of each data frame.
[0039] S2. Independently cluster all data frames in the spectral time-domain feature space and energy-domain feature space respectively, generate spectral time-domain cluster labels and energy-domain cluster labels for each data frame and pair them into composite cluster labels; for each composite cluster label, collect the initial combination features of its subordinate data frames and calculate the shared local covariance matrix.
[0040] Since the evolution of wear state is not uniformly distributed throughout the feature space, the acoustic features of different wear stages exhibit different aggregation patterns in the spectral time domain and energy domain. Therefore, this embodiment performs clustering independently in the two feature spaces to capture the local distribution pattern of data from two complementary perspectives: spectral structure and instantaneous energy.
[0041] The 16-dimensional spectral temporal feature vectors of all data frames are aggregated into one. The feature matrix of , where Let be the total number of data frames. A density-based noise-applied spatial clustering algorithm is applied to this feature matrix to perform clustering, generating an integer-form spectral temporal cluster label for each data frame. The density-based noise-applied spatial clustering algorithm can automatically discover clusters of arbitrary shapes and label samples in low-density regions as noise; its principle is well-known in the field.
[0042] The 3D energy domain eigenvectors of all data frames are aggregated into one. The feature matrix is used to perform DBSCAN clustering independently again, using another set of independently determined parameters to generate an energy domain cluster label for each data frame.
[0043] The two core parameters of the DBSCAN algorithm are the neighborhood radius. and minimum sample size . The determination of the method adopted - Distance graph method: Calculate the distance from each data point to the nearest neighbor. The distance between the nearest neighbors of all data points -Draw after sorting distance values from smallest to largest - Distance plot, the horizontal axis is the index of the sorted data points, and the vertical axis is the corresponding distance. - The distance value, the ordinate value corresponding to the inflection point where the slope of the curve changes most sharply, is selected as... ;in The value of and Consistency, to ensure The density reachability reflected by the distance map is consistent with the subsequent clustering decision criteria. In engineering practice of similar acoustic signal clustering systems, The value is typically set between 15 and 30; this embodiment follows this empirical value. The value is 20, but in other implementations it can be adjusted within the range of [10, 30] based on the total number of data frames. The spectral time-domain feature space and the energy-domain feature space have different dimensions and data distributions; the two clustering processes are performed separately. -Distance map determined independently The result The values are usually different.
[0044] For the Each data frame will have its spectral time-domain cluster label denoted as... The energy domain cluster label is denoted as Pairing two tags together forms a composite cluster tag for a data frame. For example, if the first Spectral temporal cluster labels of data frames 3 and energy domain cluster label If the value is 5, then the composite cluster label The ordered pair is (3, 5).
[0045] Reference Figure 2 Different shapes of scatter points correspond to different composite clusters, and the samples exhibit a separable structure in two dimensions, verifying the rationality of dual-domain clustering and composite cluster label division.
[0046] Because clustering in a single feature space can only reflect the density distribution of data within that space, while composite cluster labels cross-encode the clustering information of two spaces, making data frames belonging to the same composite cluster label similar in spectral temporal structure and energy response patterns, thus constructing a more refined local data grouping than single-space clustering.
[0047] For each unique composite cluster label, the initial 19-dimensional combined feature vectors of all data frames belonging to that composite cluster label are aggregated, and a shared local covariance matrix is calculated. Assume that a unique composite cluster label aggregates... There are 3 data frames, and the initial combined feature vector of each data frame is denoted as . , Calculate the mean vector of this cluster. : ; Local covariance matrix The relation is: ; In the formula, To belong to the first label under this composite cluster The 19-dimensional initial combined feature vector of each data frame is obtained by concatenating the spectral time domain feature vector and the energy domain feature vector; This is the mean vector of all initial combined feature vectors under the label of this composite cluster; This is the difference vector between the initial combined feature vector and the mean vector; This is the transpose of the difference vector; For regularization parameters; It is an identity matrix.
[0048] Regularization parameters The introduction of this is to ensure the local covariance matrix Positive definiteness and numerical stability. When the number of data frames belonging to a certain composite cluster label... When there are few or high linear correlations among the initial combination of eigenvectors, the covariance matrix without regularization may have eigenvalues of zero or close to zero, preventing the matrix inversion and logarithm operations required for calculating the Riemann distance from being performed. Regularization parameter The theoretical lower bound is determined by the smallest non-zero eigenvalue of the local covariance matrix. If it is too small, the regularization effect is insufficient; if it is too large, the covariance structure is overly smoothed. In practical engineering... Pick Reserve a margin; the selectable range is
[10] . -8 10 -4 ].
[0049] The above process is repeated for all unique composite cluster labels, calculating and storing a unique composite cluster label for each label. The local covariance matrices are used to encode the characteristic association structure between data points within the corresponding composite cluster, providing a geometric basis for subsequent calculations of the Riemann distance.
[0050] S3. Construct a sample relationship graph with data frames as nodes. For any two data frames, calculate the Euclidean distance between their initial combined features and the Riemann distance between the local covariance matrices corresponding to the labels of their composite clusters. Input the Euclidean distance and the Riemann distance into the kernel function and fuse them to obtain the edge weights between them. Combine the calibration samples with known wear state labels on the sample relationship graph, determine the wear state of all data frames through the label propagation algorithm, and output the wear test results of the engine alloy parts.
[0051] With all Each data frame constructs a fully connected sample relationship graph for the nodes, where each data frame is indexed globally. ( Unique identifier; any two data frames in the diagram and data frames There exists a weighted edge, and the edge weight is... The calculation incorporates two complementary distance metrics.
[0052] In actual engineering deployments, when the total number of data frames... When the size is large, to balance computational efficiency and memory overhead, those skilled in the art can replace the above fully connected sample relationship graph with... Nearest neighbor graph or - Sparse graph structures such as neighborhood graphs retain only the edges between each node and its nearest neighbors. Since the Gaussian kernel function assigns weights close to zero to distant sample pairs, the impact of sparsification on the final wear state recognition accuracy is negligible. This embodiment does not limit the sparsification strategy of the graph.
[0053] For any two data frames and Extract the two separately. Calculate the Euclidean distance between the initial combined feature vectors: ; In the formula, For data frames and data frames The initial combined feature Euclidean distance between them and Data frames and data frames corresponding The initial combined feature vector is identified by a global frame index, unlike the composite cluster internal index used in the calculation of the local covariance matrix. Initial combined feature vector of the identifier ; for Norm.
[0054] Get data frames and data frames Composite cluster tags and Find the composite cluster label from the already calculated and stored results. and Their respective local covariance matrices and For the local covariance matrix and The Riemann distance is calculated using the affine-invariant Riemannian metric. The affine-invariant Riemannian metric is defined on a symmetric positive definite matrix manifold, and its relation is: ; In the formula, Local covariance matrix and The Riemann distance between them; and Data frames and data frames The local covariance matrix corresponding to the label of the composite cluster, here is the index. , Use the global index of the data frame; for The square root reciprocal matrix, through the... Obtained by eigenvalue decomposition; This refers to matrix logarithm operations. It is the Frobenius norm.
[0055] The specific calculation process is as follows: Perform eigenvalue decomposition ,in The matrix consists of orthogonal eigenvectors. If is an eigenvalue diagonal matrix, then , Each diagonal element is The corresponding eigenvalue is the negative 1 / 2 power. Since the regularization parameter has already been applied... This ensures the positive definiteness of the local covariance matrix. All diagonal elements are positive. The calculation will not involve division by zero.
[0056] In practical engineering, Euclidean distance It measures the direct difference between two data frames at the eigenvalue level, while the Riemann distance... It measures the difference in covariance structure between the local data clusters to which two data frames belong, i.e., the difference in the geometric shape of the data distribution. Euclidean distance is sensitive to translation in the feature space, while Riemann distance is sensitive to rotation and scaling in the feature space. Combining the two can simultaneously capture the positional and structural differences between samples.
[0057] Edge weights are calculated using a Gaussian kernel function, the formula for which is: ; In the formula, For data frames and data frames Edge weights between them; The initial Euclidean distance between combined features; The Riemann distance between local covariance matrices; and The bandwidth parameter is a positive value.
[0058] Control the decay rate of the edge weights across the Euclidean distance. Control the decay rate of the edge weights in the Riemann distance. When or As the exponent increases, the value of the exponent term approaches negative infinity. Approaching zero means that the edge weight between data frame pairs that are farther apart is smaller, which is in line with the technical objective that the closer the data frame pairs are, the stronger the association. and The ratio of these two distances determines their relative importance in the edge weights. Much larger When the distance is equal to 0, the edge weight is mainly dominated by the Euclidean distance; otherwise, it is dominated by the Riemann distance. If the value is too large, only the nearest neighbor data frames will retain valid connections, resulting in a decrease in the connectivity of the graph. If the weight is too small, even distant sample pairs will maintain a high weight, which weakens the ability to distinguish local structures. The consequences of being too large or too small are similar. In this embodiment... Values , Values Both have a selectable value range of [missing value]. In other embodiments, adjustments can be made by those skilled in the art through cross-validation based on the characteristic distribution of actual data, within the aforementioned range.
[0059] Completed weight matrix Then, it enters the semi-supervised learning stage, where the wear status of all data frames is determined through the label propagation algorithm.
[0060] Calculate the diagonal matrix diagonal matrix diagonal elements That is, the weight matrix No. The sum of all elements in a row, with all other elements equal to 0. Calculate the probability transition matrix. probability transition matrix equal to a diagonal matrix The inverse matrix multiplied by the weight matrix ,Right now Probability transition matrix The sum of each row is 1. Characterizing tag information from data frames propagation to data frames The probability of.
[0061] Initialize a Tag matrix ,in This represents the total number of wear condition categories. For a preset number of calibration samples with known wear condition labels, in The row corresponding to the calibrated sample is set as a one-hot encoded vector with known wear state, and the row vectors corresponding to all uncalibrated samples are initially all zeros.
[0062] It should be noted that calibration samples refer to data frames whose wear condition category has been pre-confirmed through manual inspection or other known reliable means. The number of calibration samples usually accounts for a small proportion of all data frames. In this embodiment, the wear condition is divided into three categories: normal, slight wear, and severe wear. The value is 3, and the calibration samples account for one percent of all data frames. In other implementations, the total number of wear condition categories is... The proportion of calibrated samples can be adjusted within the range of one to ten percent depending on the actual diagnostic scenario.
[0063] By iteratively updating the label matrix, the label matrix at the next time step is equal to the probability transition matrix. Multiply by the label matrix at the current time, i.e. After each iteration, a label clamping operation is performed, forcibly resetting the labels of the corresponding rows of the labeled samples to their initial values in a known state, in order to prevent the labels of the labeled samples from being overwritten by information from the surrounding unlabeled samples during the iteration process.
[0064] Check the convergence condition; if it is satisfied... Then the iteration stops, where The convergence threshold is determined by those skilled in the art through conventional experiments based on the required computational accuracy. In this embodiment... for .
[0065] For any unlabeled data frame The wear state is determined by the converged tag matrix. The Middle The category corresponding to the element with the largest row value, i.e. This allows for the determination of the wear status of all data frames for the engine alloy components.
[0066] Typical parameters from an engine wear acoustic dataset were used for verification. The dataset includes three wear states: normal, slight wear, and severe wear, totaling 30,000 data frames. All semi-supervised methods used only 1% of the labeled samples. Evaluation metrics included accuracy, precision, recall, and [missing data] for wear state identification. Fraction.
[0067] The comparative schemes include: the complete scheme of this invention; the Euclidean distance-only scheme, which uses only the Euclidean distance between the initial combined features and removes the Riemann distance part when constructing the sample relationship graph; the spectral time-domain feature-only scheme, which does not include energy domain features and only uses 16-dimensional spectral time-domain features; and the support vector machine scheme, which uses the complete 19-dimensional feature set for supervised learning training.
[0068] Reference Figure 3 The horizontal axis represents the evaluation metrics, and the vertical axis represents the scores of each solution on the corresponding metrics. This invention performed best on all evaluation metrics, with an accuracy of 97.5%, a precision of 96.8%, and a recall of 97.1%. The score was 96.9%. The accuracy of the Euclidean distance-only scheme was 93.2%, and the precision, recall, and... The scores were 92.5%, 92.8%, and 92.6%, respectively. The accuracy of the spectral time-domain feature-only scheme was 91.5%, and the precision, recall, and... The scores were 90.7%, 91%, and 90.8%, respectively. The accuracy of the support vector machine approach was 88.3%, and the precision, recall, and... The scores were all around 88%.
[0069] Compared to the Euclidean distance-only approach, this invention improves accuracy by 4.3 percentage points. This is because Riemann distance can perceive differences in the covariance structure of data clusters and is more sensitive to rotation and scaling changes in the feature space, thereby improving the accuracy of edge weights in the sample relationship graph in representing differences in wear states. Compared to the spectral time-domain feature-only approach, this invention improves accuracy by 6 percentage points. Energy domain features are particularly adept at capturing non-stationary transient impact components caused by wear faults, serving as a powerful complement to spectral time-domain features. Compared to the support vector machine approach, this invention achieves 9.2 percentage points higher accuracy. The label propagation algorithm fully utilizes the data distribution information contained in a large number of unlabeled samples, demonstrating a significant advantage in scenarios where labeled samples are scarce.
[0070] This invention also discloses an acoustic emission-based engine alloy component wear testing system, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the acoustic emission-based engine alloy component wear testing method according to this invention.
[0071] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.
[0072] In the description of this specification, "multiple" or "several" means at least two, such as two, three or more, unless otherwise expressly and specifically defined.
Claims
1. A method for testing wear of engine alloy components based on acoustic emission, characterized in that, include: The system acquires continuous sound signals from the engine alloy components during operation and divides these signals into multiple data frames. The sound signals are acoustic emission signals acquired by acoustic emission sensors. For each data frame, a spectral time-domain feature set is extracted. The absolute value of the Teager-Kaiser energy operator is applied to the data frame to obtain an instantaneous energy sequence, and an energy domain feature set is extracted. Z-score normalization is performed on each dimension of the spectral time-domain feature set and the energy domain feature set. The normalized spectral time-domain feature set and the energy domain feature set are concatenated to form an initial combined feature set. All data frames are independently clustered in the spectral time-domain feature space and the energy-domain feature space, respectively, and spectral time-domain cluster labels and energy-domain cluster labels are generated for each data frame and paired into composite cluster labels. For each composite cluster label, the initial combined features of its subordinate data frames are collected to calculate the shared local covariance matrix. A sample relationship graph is constructed with data frames as nodes. For any two data frames, the Euclidean distance between their initial combined features and the Riemann distance between the local covariance matrices corresponding to their respective composite cluster labels are calculated. The Euclidean distance and the Riemann distance are input into a kernel function and fused to obtain the edge weights between them. The wear state of all data frames is determined by combining the calibration samples with known wear state labels on the sample relationship graph and using a label propagation algorithm. The wear test results of the engine alloy components are then output.
2. The method for testing wear of engine alloy components based on acoustic emission according to claim 1, characterized in that, The step of dividing the acquired audio signal into multiple data frames includes: setting the frame length of the data frame to 2048 sampling points, the frame shift to 1024 sampling points to form a 50% overlap rate, and applying a Hamming window to each data frame to reduce spectral leakage.
3. The method for testing wear of engine alloy components based on acoustic emission according to claim 1, characterized in that, The extraction of the spectral time-domain feature set for each data frame includes: For each data frame, the 13-dimensional Mel frequency cepstral coefficients, spectral centroid, spectral roll-off point, and zero-crossing rate are extracted, and these four features are combined to form a spectral time-domain feature set.
4. The method for testing wear of engine alloy components based on acoustic emission according to claim 1, characterized in that, The process of applying the Teager-Kaiser energy operator to the data frame, taking the absolute value to obtain the instantaneous energy sequence, and extracting the energy domain feature set includes: The mean, standard deviation, and energy entropy of the instantaneous energy sequence within each data frame are calculated, and these three indicators are used as the energy domain feature set.
5. The method for testing wear of engine alloy components based on acoustic emission according to claim 1, characterized in that, The clustering algorithm is a density-based noise-based spatial clustering algorithm; during independent clustering, for each data point, the distance to its next node is calculated. The distances to the nearest neighbors are used to plot a k-distance graph, where all data points are sorted in ascending order of distance values. The distance value corresponding to the inflection point of the curve is selected as the neighborhood radius. ; Preset minimum sample size Based on the neighborhood radius and the minimum number of samples Perform clustering; the Values and Consistent.
6. The method for testing wear of engine alloy components based on acoustic emission according to claim 1, characterized in that, The calculation of the shared local covariance matrix includes: For those belonging to the same composite cluster label Initial combined feature vectors ( ), Calculate the mean vector: ； Local covariance matrix The relation is: ； In the formula, To belong to the first label under this composite cluster The initial combined feature vector of each data frame, This is the difference vector between the initial combined feature vector and the mean vector. Difference vector transpose, For regularization parameters, It is an identity matrix.
7. The method for testing wear of engine alloy components based on acoustic emission according to claim 1, characterized in that, For any two data frames, the Euclidean distance between the initial combined features of the two data frames and the Riemann distance between the local covariance matrices corresponding to the labels of the composite clusters to which the two data frames belong are calculated. The Euclidean distance and the Riemann distance are then input into a kernel function and fused to obtain the edge weights between the two data frames, including: For any two data frames, calculate the Euclidean distance between the initial combined features; For the local covariance matrix corresponding to the composite cluster labels of the two data frames and The Riemann distance is calculated using the affine-invariant Riemann metric. Edge weights are calculated using a Gaussian kernel function.
8. The method for testing wear of engine alloy components based on acoustic emission according to claim 1, characterized in that, The step of determining the wear status of all data frames using the tag propagation algorithm includes: Construct the probability transition matrix probability transition matrix equal to a diagonal matrix inverse matrix multiplied by weight matrix Weight matrix The weight matrix and diagonal matrix of the sample relationship graph. The diagonal elements are the weight matrix. The sum of all elements in the corresponding row; Initialize the label matrix The row vectors corresponding to the calibrated samples are set as one-hot encoded vectors with known wear states, while the row vectors corresponding to the uncalibrated samples are initially set to all zeros. By iteratively updating the label matrix, the label matrix at the next time step is equal to the probability transition matrix. Multiply by the label matrix at the current time step, and reset the labels of the labeled samples to the known state after each iteration, until the label matrix is fully updated. Convergence determines the wear status of all data frames.
9. The method for testing wear of engine alloy components based on acoustic emission according to claim 7, characterized in that, The relationship of the Gaussian kernel function is: ； In the formula, For data frames and data frames Edge weights between them The Euclidean distance between the initial combined features. The Riemann distance between local covariance matrices. and The bandwidth parameter is a positive value.
10. A wear testing system for engine alloy components based on acoustic emission, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the acoustic emission-based wear testing method for engine alloy components according to any one of claims 1-9.