Rotary machinery health index self-adaptive construction method and system and medium

Abstract

The invention provides a rotary machine health index self-adaptive construction method and system and a medium, and belongs to the technical field of equipment maintenance, and the method comprises the steps: extracting the time domain, frequency domain and time-frequency domain features of an equipment vibration signal, including complexity entropy and relative similarity, and carrying out the two-dimensional image conversion, gray features, gray difference statistical features and gray co-occurrence matrix features are extracted, and a feature set is constructed; according to correlation, monotonicity, robustness and stability index sorting features, carrying out line-by-line displacement and coupling to construct a mixed feature arrangement sample; according to a feature extraction mode, the feature set is divided into multiple classes, a feature relation graph is constructed, graph nodes represent feature types and values, and edges represent information relations; extracting hidden layer spatio-temporal features by using a Transform-based spatio-temporal fusion convolution auto-encoder, and extracting graph structure features by using a GCN-based graph convolution auto-encoder; the spatial-temporal features and the decoder reconstruction features are fused through a bidirectional cross attention auto-encoder, and finally the health indexes of the equipment are output through quadratic function constraints. The method can deeply mine the degradation information between the features and improve the overall performance of the health indexes.

Description

Technical Field

[0001] This invention relates to the field of equipment maintenance technology, specifically to an adaptive construction method, system, and medium for health indicators of rotating machinery. Background Technology

[0002] Rotating machinery is a crucial piece of equipment in modern industrial production, widely used in aerospace, shipbuilding, energy, chemical, and transportation industries. As rotating machinery develops towards larger, more complex, more precise, and more intelligent structures, harsh operating environments and complex working conditions increase the probability of performance degradation, failure, and malfunctions. To address the challenges of safe operation and efficient maintenance of rotating machinery, Prognostic and Health Management (PHM) technology provides a more scientific strategy for predictive maintenance and operation management. Health Indicator (HI), as a vital component of PHM, can accurately grasp the operating status of rotating machinery, identify potential faults, and provide a basis for proactive countermeasures. Therefore, research on HI construction methods for rotating machinery is of great significance for improving the reliability of rotating machinery operation.

[0003] By constructing the HI index through deep feature fusion, the changing patterns of equipment health status can be quantified. Traditional methods mainly include Bayesian models, Markov models, DS evidence theory, SVM, similarity fusion, etc. With the development of deep learning, CNN, LSTM, autoencoders and other methods have been gradually applied, which have effectively improved the fusion accuracy and reduced the difficulty of constructing the HI index when the data volume is large and the data relationship is complex. However, the existing technologies have the following problems: (1) Existing feature extraction and selection methods mostly extract features from the time domain, frequency domain and time-frequency domain of the signal, which makes it difficult to completely and effectively extract representative and degradation-sensitive features, and it is difficult to describe the dynamic process of rotating machinery performance degradation. (2) Existing HI construction networks are limited to simple inputs in the form of feature splicing into one-dimensional vectors, and do not effectively represent the mutual coupling relationship between features, making it difficult to deeply explore the degradation information between features. (3) Existing HI construction methods do not deeply discuss the correlation features of signal features in the spatial dimension, and the method based on a simple network structure is difficult to fully extract the deep degradation features of the data, making it difficult to express the subtle changes in health status during the mechanical degradation process. Summary of the Invention

[0004] This invention provides an adaptive construction method for health indicators of rotating machinery. This method can deeply mine the degradation information between features and improve the overall performance of health indicators.

[0005] An adaptive method for constructing health indicators for rotating machinery includes the following steps: The time-domain, frequency-frequency-domain, complexity entropy, and relative similarity features of the equipment vibration signal are extracted by variational mode decomposition. The equipment vibration signal is then transformed into two-dimensional images in the time domain, frequency domain, and time-frequency domain to extract gray-level features, gray-level difference statistical features, and gray-level co-occurrence matrix features, and a feature set is constructed. The feature selection method based on correlation, monotonicity and robustness indicators is used to sort the features according to their scores. The sorted features are then shifted row by row and coupled in a way that constructs a two-dimensional hybrid feature arrangement coupled sample. The feature set is divided into multiple categories based on the feature extraction method, and a feature relationship graph is constructed. The graph nodes represent feature type, name, and feature value, while the edges represent similar information, related information, and mutually related information. A pre-trained Transformer-based spatiotemporal fusion convolutional autoencoder is used to extract the hidden spatiotemporal features of the hybrid feature arrangement coupled samples; a pre-trained GCN-based graph convolutional autoencoder is used to extract the hidden graph structure features of the feature relationship graph. The encoder of the bidirectional cross-attention autoencoder fuses the hidden spatiotemporal features to obtain fused features, which are then reconstructed by the decoder to obtain reconstructed features. Finally, the corresponding health indicators are output through quadratic function constraints.

[0006] Preferably, the encoder part of the spatiotemporal fusion convolutional autoencoder includes a spatial feature extraction module, a Transformer encoder, and a deep feature extraction module; the decoder part adopts an asymmetric decoder, including a single deconvolution layer and an upsampling layer; the spatial feature extraction module includes a two-dimensional convolutional layer, a pooling layer, and a reshape layer connected in sequence, and uses a BN layer and a ReLU activation function for parameter tuning; the deep feature extraction module includes a convolutional layer, a Conv layer, and a pooling layer connected in sequence. The spatial feature extraction module extracts the spatial features of the mixed feature arrangement coupled samples. The Transformer encoder uses a multi-head attention mechanism to perform global spatiotemporal feature extraction on the spatial features. The deep feature extraction module expands the global spatiotemporal features to obtain the hidden spatiotemporal features after spatiotemporal fusion. The encoder part of the graph convolutional autoencoder consists of two GCN layers, and the decoder part consists of two fully connected layers. The encoder's first GCN layer aggregates the first-order neighbor information of each feature node in the feature relationship graph and performs a nonlinear transformation; the second GCN layer generates a low-dimensional vector representation of each feature node in the feature relationship graph to obtain the hidden layer graph structure features; the decoder reconstructs the original node feature matrix based on the hidden layer graph structure features through a two-layer fully connected network, and recovers the original feature values ​​from the relationship embedding. The bidirectional cross-attention autoencoder is a dual-channel encoder structure based on the bidirectional cross-attention mechanism and is connected to an adaptive weight vector module. The dual-channel encoder structure maps and fuses the attention mechanisms between the hidden layer spatiotemporal features and the hidden layer graph structure features to obtain fused features; the adaptive weight vector module dynamically selects the fused features. A health metric construction model is constructed based on the spatiotemporal fusion convolutional autoencoder, graph convolutional autoencoder, and bidirectional cross-attention autoencoder.

[0007] Preferably, the health indicator construction model is obtained by training the training set, wherein the total loss of the model is: ; In the formula, As weight, Loss 1 represents the reconstruction loss of the spatiotemporal fusion convolutional autoencoder; where Loss 2 represents the reconstruction loss of the graph convolutional autoencoder; Loss 3 represents the feature fusion and reconstruction loss in the bidirectional cross-attention autoencoder, ensuring that the fused features can be correctly reconstructed and that information integrity is maintained; Loss 4 represents the output loss for health indicators; all use the Huber loss function: In the formula, δ For error hyperparameters, when δ When ~0, Huber loss tends to mean absolute error; when δ As the value approaches infinity, Huber loss tends to be the mean squared error.

[0008] Preferably, it also includes introducing multi-scale Savitzky-Golay filtering to eliminate spikes in the HI curve and smooth the HI curve.

[0009] Preferably, the time-domain features include mean, absolute mean, variance, standard deviation, root amplitude, root mean square value, peak value, peak-to-peak value, maximum value and minimum value, as well as dimensionless features; the dimensionless features include waveform index, peak index, impulse index, margin index, skewness index and kurtosis index; The time-frequency domain features are obtained by performing three-level wavelet packet decomposition on the equipment vibration signal to obtain eight continuous wavelet energy spectrum features; The complexity entropy features include features extracted from time-domain information entropy, frequency-domain information entropy, permutation entropy, sample entropy, approximation entropy, energy entropy, and scatter entropy. The relative similarity features include 5 time-domain RS features and 5 frequency-domain RS features, namely Euclidean distance, Manhattan distance, Chebyshev distance, cosine distance and Pearson distance.

[0010] Preferably, the step of converting the device vibration signal into a time-domain, frequency-domain, and time-frequency-domain two-dimensional image includes the following steps: The sample amplitudes of the time-domain vibration signal and the frequency-domain vibration signal are normalized to [0,1] respectively. Then, the normalized sample amplitudes are converted into image pixel values ​​to construct a two-dimensional image of the vibration signal in the time domain and a two-dimensional image of the frequency domain. Consider analyzing vibration signals , Using the Dirac function (which, in the time domain, can be considered an ideal model of a signal with transient characteristics), the vibration signal is transformed from a one-dimensional time signal into a two-dimensional time-frequency signal through short-time Fourier transform: ; In the formula, A Instantaneous amplitude, t For time, t 0 For a certain time and location, It is a Gaussian window function. For frequency; because The energy distribution of the short-time Fourier transform of the Dirac function is expressed as: ; Calculate the frequency derivative of the short-time Fourier transform: ; For any , making Two-dimensional Group delay is calculated as follows: ; Introducing the transient extraction operator TEO: ; in: ; In the formula, Δ represents the time support range of the window function. Therefore, the transient extraction transform is expressed as: ; Due to Dirac Functions have the following properties: ; The above transient extraction transform can then be expressed as: ; The transient features of the extracted transformation are converted to grayscale to obtain a two-dimensional image in the time-frequency domain.

[0011] Preferably, the gray-level difference statistical features include gray-level mean, gray-level variance, gray-level difference statistical mean, gray-level difference statistical contrast, gray-level difference statistical image entropy, maximum and minimum pixel values ​​within the region of interest, pixel value range, pixel value mean absolute deviation, and pixel value root mean square. The gray-level co-occurrence matrix features are taken from the gray-level co-occurrence matrices in four directions: 0°, 45°, 90° and 135°. The energy, entropy and moment of inertia of the gray-level co-occurrence matrix in the four directions are extracted.

[0012] Preferably, the feature selection method based on correlation, monotonicity, and robustness indices for selecting the optimal feature includes: ; In the formula, C is the final feature score. As a correlation indicator, As a monotonicity indicator, For robustness indicators, Sta As a stability index, the four feature values ​​are in the range of [0,1]. The larger the value, the higher the feature score, indicating that the feature is more representative of the degradation trend of the device.

[0013] This invention also proposes an adaptive construction system for health indicators of rotating machinery, the system comprising: The feature processing module is used to extract time-domain, frequency, time-frequency domain, complexity entropy, and relative similarity features of equipment vibration signals through variational mode decomposition. It then performs time-domain, frequency-domain, and time-frequency domain two-dimensional image transformations on the equipment vibration signals, extracting gray-level features, gray-level difference statistical features, and gray-level co-occurrence matrix features to construct a feature set. Furthermore, it sorts the features based on scores using feature selection methods based on correlation, monotonicity, robustness, and stability indices. The sorted features are then shifted row by row and coupled to construct a two-dimensional hybrid feature arrangement coupled sample. Based on the feature extraction method, the feature set is divided into multiple feature classes, and a feature relationship graph is constructed. In this graph, nodes represent feature type, name, and feature value, while edges represent similar information, related information, and mutually related information. The health indicator construction module is used to extract the hidden spatiotemporal features of the mixed feature arrangement coupled samples by an encoder based on a pre-trained Transformer-based spatiotemporal fusion convolutional autoencoder; extract the hidden graph structure features of the feature relationship graph by an encoder based on a pre-trained GCN-based graph convolutional autoencoder; fuse the hidden spatiotemporal features with the hidden spatiotemporal features by an encoder based on a bidirectional cross-attention autoencoder to obtain fused features, reconstruct the features by a decoder, and output the corresponding health indicators by quadratic function constraints.

[0014] The present invention also proposes a computer-readable storage medium storing a data processing program, wherein the steps of an adaptive construction method for health indicators of rotating machinery are implemented when the data processing program is executed by a processor.

[0015] The beneficial effects of this invention are: This invention proposes an adaptive construction method for health indicators of rotating machinery. This method proposes a multi-dimensional perspective vibration signal feature extraction method, designs a multi-modal signal multi-domain hybrid feature extraction method, and proposes a two-dimensional image conversion method for the time domain, frequency domain, and time-frequency domain, as well as an image feature extraction method, enhancing the completeness and effectiveness of features. The method establishes a graph representation method based on permutation coupling rules to enrich the interrelationships between features, proposing a feature representation learning relationship graph that considers feature type, name, feature value, and feature node correlation, enabling the representation and mining of complex degradation relationships between different features. The method establishes a shape-constrained Hierarchical Indicator (HI) construction model integrating graph convolutional networks, Transformers, and a double cross-attention mechanism, extracting feature topology and spatiotemporal relationships to complete automatic HI construction. A novel multi-constraint model structure is designed, and the overall performance of HI is improved based on minimizing the loss function. Attached Figure Description

[0016] Figure 1 This is a flowchart of the adaptive construction method for health indicators of rotating machinery according to an embodiment of the present invention; Figure 2 This is an overall model architecture diagram of the adaptive construction method for health indicators of rotating machinery according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the two-dimensional image conversion of vibration signals according to an embodiment of the present invention; Figure 4 This is a schematic diagram of vibration signal frequency domain image conversion according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the time-frequency domain two-dimensional image conversion of vibration signals according to an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the specific steps of multidimensional hybrid feature arrangement coupling in an embodiment of the present invention; Figure 7 This is a schematic diagram of the feature map construction result according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the overall architecture of the health indicator construction model TGBCAE according to an embodiment of the present invention; Figure 9 This is the Transformer model structure according to an embodiment of the present invention; Figure 10 This is a structural diagram of the bidirectional cross-attention mechanism autoencoder model according to an embodiment of the present invention; Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Example 1 Currently, there is no existing method for automatically constructing Health Indicators (HIs) that integrates multi-dimensional hybrid features, feature map representation, and graph convolutional autoencoders. This invention proposes an adaptive construction method for rotating machinery health indicators based on multi-dimensional hybrid feature map representation and multi-constraint autoencoder networks. The process steps are as follows: Figure 1 As shown, the method for constructing health indicators for rotating machinery using multidimensional hybrid feature map representation and multi-constraint autoencoder networks is as follows: Figure 2 As shown, it mainly includes multi-dimensional feature extraction, feature map representation establishment, HI model construction, and HI spurious removal. It includes the following steps:

[0019] S1: Extract the time-domain, frequency, time-frequency domain, complexity entropy, and relative similarity features of the equipment vibration signal through variational mode decomposition. Then, perform time-domain, frequency-domain, and time-frequency domain two-dimensional image transformation on the equipment vibration signal to extract gray-level features, gray-level difference statistical features, and gray-level co-occurrence matrix features, and construct a feature set.

[0020] S2: The features are sorted according to their scores using a feature selection method based on correlation, monotonicity, robustness, and stability indices. The sorted features are then shifted row by row and coupled in a way that constructs a two-dimensional hybrid feature arrangement coupled sample.

[0021] S3: Divide the feature set into multiple categories based on the feature extraction method and construct a feature relationship graph; where the graph nodes are feature type, name and feature value, and the edges are similar information, related information and mutually related information.

[0022] S4: The encoder uses a pre-trained Transformer-based spatiotemporal fusion convolutional autoencoder to extract the hidden spatiotemporal features of the hybrid feature arrangement coupled samples; the encoder uses a pre-trained GCN-based graph convolutional autoencoder to extract the hidden graph structure features of the feature relationship graph.

[0023] S5: The encoder of the bidirectional cross-attention autoencoder fuses the hidden spatiotemporal features to obtain the fused features, which are then reconstructed by the decoder to obtain the reconstructed features. Finally, the corresponding health indicators are output through quadratic function constraints.

[0024] Among them, Variational Mode Decomposition (VMD) in S1 extracts the time-domain, frequency, time-frequency domain, complexity entropy, and relative similarity features of the equipment vibration signal, ensuring the completeness and effectiveness of the equipment information and improving the ability to characterize the nonlinear behavior of the system. Specifically, this includes: (1) Time-domain characteristics, including mean T 1 Absolute mean T 2 ,variance T 3 Standard deviation T 4 Root amplitude T 5 Root mean square value T 6 Peak T 7 Peak-to-peak value T 8 Maximum value T 9 and minimum value T 10 Dimensionless characteristics mainly include waveform indices. T 11 Peak indicators T 12 Pulse Indicators T 13 Margin indicators T 14 Skewness index T 15 kurtosis index T 16 .

[0025] (2) Frequency domain characteristics, including frequency domain characteristics derived from F 1 - F 13 This indicates that the internal properties of a system can be identified based on the magnitude of vibration energy, the location of the main frequency band, and the degree of dispersion or concentration of the spectrum reflected by the characteristics.

[0026] (3) Time-frequency domain features: The signal is decomposed into three layers of wavelet packets to obtain eight continuous wavelet energy spectrum features, which are respectively derived from... TF 1 -TF 8 It can perform multi-scale refined analysis on stationary or non-stationary signals, effectively extracting the nonlinear characteristics of signals under weak and strong noise interference.

[0027] (4) Complexity entropy, which quantifies the dynamic changes of a system and analyzes signal distribution characteristics and complexity. This invention utilizes time-domain information entropy. E 1Frequency domain information entropy E 2 Permutation entropy E 3 Sample entropy E 4 Approximate entropy E 5 Energy entropy E 6 Dispersion entropy E 7 Extract features.

[0028] (5) Relative similarity ( RS This invention extracts five time-domain features. RS feature( RS 1 , RS 2 , RS 3 , RS 4 , RS 5 ) and 5 frequency domains RS feature( RS 1 , RS 2 , RS 3 , RS 4 , RS 5 The distances are Euclidean distance, Manhattan distance, Chebyshev distance, cosine distance, and Pearson distance, respectively. (Time domain) RS These features can promptly reflect the overall fluctuations of a signal, reduce noise interference, and avoid significant fluctuations caused by sudden signal changes or anomalies. Frequency domain RS features have advantages in noise reduction, clutter removal, and classification comparison.

[0029] This embodiment extracts a total of 54 multi-domain features. If the number of sub-modes after signal decomposition is n, then the total number of features is... indivual.

[0030] The two-dimensional image conversion in S1 includes the following steps: 1) Time-domain two-dimensional image conversion of vibration signals Existing methods primarily extract one-dimensional vibration features in the time, frequency, and time-frequency domains. However, one-dimensional vibration does not consider the relationships between non-adjacent sample points, making it difficult to extract hidden features. Therefore, this paper introduces a method to reconstruct a one-dimensional vibration signal into a grayscale image, extracting its feature information from an image perspective. This provides a method for exploring the two-dimensional features of the original signal, effectively extracting the relationships between non-adjacent sample points. Furthermore, this method can be calculated without any predefined parameters, reducing errors caused by experience.

[0031] A method for reconstructing a one-dimensional vibration signal into a grayscale image is introduced, extracting its feature information from the image perspective. This method can reveal the one-dimensional hidden features of the vibration signal in a two-dimensional image. Specifically, the vibration signal sample amplitude is first normalized to [0,1], and then the normalized sample amplitude is converted into image pixel values. Figure 3 This is a schematic diagram illustrating the process of converting a vibration signal into a vibration image. The sample amplitude and the corresponding image pixel value are represented by the following formula:

[0032] ; in, M Indicates the image dimensions, including length and width. j =1: M , k =1: M Therefore, the number of pixels in the vibration image is equal to the number of sample points in the vibration signal. P [ j , k ]for M × M The intensity of the corresponding pixel in the image of large and small vibrations. B [•] is the amplitude of the sample point transformation in the vibration signal, which is the normalized value of the sample point multiplied by 255.

[0033] 2) Frequency domain two-dimensional image conversion of vibration signals To address the difficulty in extracting frequency components from time-domain analysis alone, this invention constructs a vibration image based on the frequency domain of the vibration signal, building upon the construction of a time-domain vibration image. Specifically, the original signal is first frequency-domain transformed, followed by image transformation, ensuring the extraction of two-dimensional features in the frequency-domain image. The vibration signal frequency-domain image transformation process is as follows: Figure 4 As shown.

[0034] 3) Time-frequency domain two-dimensional image conversion of vibration signals To address the challenges of time-frequency analysis and transient feature extraction for non-stationary signals, this invention proposes a two-dimensional time-frequency domain image construction method based on transient extraction and grayscale transformation. This method extracts transient components from complex signals, effectively identifies transient events in signals, and enhances resistance to background noise.

[0035] Transient extraction transform is a post-processing method of short-time Fourier transform, which first considers the analysis of impulse signals. , Using the Dirac function (which, in the time domain, can be considered an ideal model of a signal with transient characteristics), the vibration signal is transformed from a one-dimensional time signal into a two-dimensional time-frequency signal through short-time Fourier transform: ; In the formula, A is the instantaneous amplitude. t For time, t 0 For a certain time and location, It is a Gaussian window function. For frequency; because The energy distribution of the short-time Fourier transform of the Dirac function is expressed as: ; Because the window function is time-compact, its short-time Fourier transform energy reaches its maximum at time t = t0. The short-time Fourier transform of the Dirac delta function introduces a group delay. To accurately estimate the group delay of each Dirac function, we first calculate the frequency derivative of its short-time Fourier transform, as shown in the following formula:

[0036] ; For any , making Two-dimensional Group delay can be calculated as: ; To remove aliased time-frequency coefficients and retain only those at time t0, the Transient-extracting Operator (TEO) was proposed: ; in: ; In the formula: Δ represents the time support range of the window function. Therefore, the transient extraction transform can be expressed as:

[0037] ; Due to Dirac Functions have the following properties: ; Therefore, the above transient extraction transform can be expressed as: ; Therefore, after transient feature extraction and transformation, irrelevant information is stripped away, highlighting the transient impact components related to the fault, and the energy of the signal will be more concentrated during time-frequency analysis.

[0038] This invention extracts and transforms transient features from vibration signals and converts them to grayscale to more intuitively display the intensity distribution of transient features while preserving their dynamic information. This provides cleaner, more prominent, and higher-quality feature input for subsequent high-intensity analysis (HI) construction. The process is as follows: Figure 5 As shown.

[0039] This invention extracts two-dimensional image features across three image domains, including gray-level features, gray-level difference statistical features, and gray-level co-occurrence matrix features. The extracted features include the gray-level mean. I 1 Gray-scale variance I 2 Gray-scale difference statistical mean I 3 Gray-scale difference statistical contrast I 4 Gray-scale difference statistical image entropy I 5 Maximum pixel value within the region of interest I 6 and minimum value I 7 Pixel value range I 8 Mean absolute deviation of pixel values I 9 Root mean square of pixel value I 10 wait.

[0040] The gray-level co-occurrence matrix (GLCM) is a function of pixel distance and angle. It reflects comprehensive information about the image in terms of direction, interval, magnitude of change, and speed by calculating the correlation between the gray levels of two points at a certain distance and direction in the image. This invention takes four directions—0°, 45°, 90°, and 135°—to obtain the corresponding GLCMs for these four directions. The energy of the GLCMs in these four directions is then extracted. I 11 - I 14 ,entropy I 15 - I 18 Moment of inertia I 19 - I22 There are a total of 12 features. Therefore, this invention extracts a total of 22 two-dimensional features from vibration signals.

[0041] Furthermore, S2 provides the optimal feature ranking method and the method for constructing coupled samples with mixed features, specifically: Degradation features should maintain a certain correlation with time, exhibit an overall increasing or decreasing trend during the degradation process, and maintain a certain degree of robustness to outliers. Therefore, this invention employs a feature selection method that integrates three indicators—correlation, monotonicity, and robustness—to select the optimal feature, as shown in the following formula:

[0042] ; In the formula, C is the final feature score. As a correlation indicator, As a monotonicity indicator, For robustness indicators, Sta As a stability index, the four feature values ​​are in the range of [0,1]. The larger the value, the higher the feature score, indicating that the feature is more representative of the degradation trend of the device.

[0043] Furthermore, this invention first sorts the extracted, disordered optimal features according to their scores, ensuring fixed regions for high-scoring and low-scoring features. Then, to extract the relationships between non-adjacent features and improve the network's ability to extract local information between features, a two-dimensional feature array is constructed by shifting features row by row and coupling them at each row. This allows for the arrangement and coupling of mixed features, ensuring usability for subsequent network operations. Specifically, as follows... Figure 6 As shown.

[0044] S3 provides a method for constructing the feature relationship graph, specifically: Nodes and node features are crucial attributes of road graphs. The meaning of a node determines the representational relationship of edges, and node features reflect the state information of the node. This invention uses different features of vibration signals as graph nodes, and the edges of the road graph represent the dependencies between these features. When constructing the graph structure, feature type and feature value are simultaneously incorporated into the node design. By explicitly distinguishing between feature semantics and numerical relationships, the graph's ability to represent complex data is enhanced. Based on the feature extraction method, features are divided into the following eight categories:

[0045] ① The “one-dimensional → primitive + modal → time domain” feature is divided into type 1, named OF; ② The "one-dimensional → original + modal → frequency domain" feature is divided into type 2, named OT; ③ The "one-dimensional → primitive + modal → time-frequency domain" feature is divided into type 3, named OTF; ④ The “One-Dimensional → Primitive + Modal → Entropy” feature is divided into type 4, named OE; ⑤ The "one-dimensional → original + modal → relative similarity" feature is divided into 5 types, named OS; ⑥ The "2D image → temporal domain" feature is divided into type 6, named TT; ⑦ The "two-dimensional image → frequency domain" feature is divided into type 7, named TF; ⑧ The “Two-dimensional image → Time-frequency domain” feature is divided into type 8, named TTF.

[0046] Therefore, each node in the constructed feature graph will include feature type, name, and feature value. By explicitly storing the type and name, feature traceability is achieved. Node attributes, as metadata containers, provide a unified expression framework for heterogeneous features from different domains (time domain, frequency domain, entropy, etc.) with different physical meanings and calculation methods, while satisfying the dual requirements of visualization and interpretability.

[0047] To capture the correlations between features from different dimensions and form a multi-layered feature relationship network, the feature relationship graph constructed in this invention includes three types of relationships: semantic layer homogeneous connections, linear numerical layer correlation coefficients, and non-linear information layer mutual information. First, semantic layer homogeneous connections prevent similar features from being segmented into isolated subgraphs due to weak numerical statistical relationships, thus losing type-level correlation information. Second, the correlation coefficients in the linear numerical layer identify potential linear dependencies between features, solving the collinearity problem. Finally, the mutual information in the non-linear information layer captures arbitrary statistical relationships, compensating for the shortcomings of correlation coefficients, reducing the impact of noise and outliers, and uncovering hidden relationships between features.

[0048] In summary, the established graph structure is as follows: Figure 7 As shown, each type of node is assigned a different color, the node number represents the specific feature under each type of feature, and the edges between nodes include information of the same type (green), related information (blue), and mutually related information (red).

[0049] S4 and S5 present the overall structure of the health metric construction model TGBCAE, including the spatiotemporal fusion convolutional autoencoder TCAE, the graph convolutional autoencoder GCAE, and the bidirectional cross-attention autoencoder. The overall architecture is as follows: Figure 8 As shown.

[0050] (1) Feature extraction of spatiotemporal fusion convolutional autoencoder To address the problems of traditional temporal feature extraction models, such as difficulty in capturing long-distance features, inability to extract spatial features, and low operating efficiency, this invention employs a portion of the Transformer structure to complete spatiotemporal feature extraction. The Transformer is a network based on a self-attention mechanism. This model can effectively handle long-term dependencies that change over time and improve operating efficiency through parallel computation. The Transformer structure is as follows: Figure 9 The model consists of an encoder module on the left and a decoder module on the right. To comprehensively consider the importance of the information contained in the input sequence, a multi-head self-attention mechanism is employed to focus on information from different representation subspaces at different positions, completing the modeling representation of different constraints and achieving deep spatial feature extraction. Before the raw data enters the encoder layer, the Transformer first passes it through a positional encoding layer, using sine and cosine positional encoding to add relative positional information, enabling the model to better capture temporal features.

[0051] The degradation data features of rotating machinery monitoring exhibit interactions and dynamic changes over time, displaying strong temporal dependence. Furthermore, the data itself is inherently time-dependent. To address the challenges of comprehensively mining degradation relationships among features under incomplete spatiotemporal conditions and the difficulty in representing multi-spatiotemporal degradation information in complex scenarios using supervised high-resolution (HI) deep learning models, a spatiotemporal fusion convolutional autoencoder is established to achieve deep representation of degradation features.

[0052] ① Spatial feature extraction: Two-dimensional convolutional layers are used to extract spatial information between features. Pooling layers are added to extract smaller and higher-level features from the feature map, which can speed up the calculation of subsequent layers and also has filtering function. Reshape layers are used to rearrange spatial features, and BN layers and ReLU activation functions are used to simplify the parameter tuning process.

[0053] ② Temporal feature extraction: The encoder part of the Transformer is used to extract global spatiotemporal features. The multi-head self-attention mechanism is used to learn information from different representation subspaces, and positional encoding enables the model to better capture temporal features.

[0054] ③ Based on the above spatiotemporal features, new convolutional layers, Conv layers, and Maxpooling layers are used to extract deeper spatiotemporal features with a higher level of abstraction. The spatiotemporal features of each channel are then expanded to obtain the high-level features after spatiotemporal fusion.

[0055] ④ An asymmetric decoder is used, which completes the model decoder through deconvolution and upsampling layers. This structure uses a single deconvolution and upsampling operation to complete sample reconstruction. BN layers and ReLU functions are used to reduce the computational cost of the model and alleviate model overfitting.

[0056] (2) Graph structure feature extraction from graph convolutional autoencoder Graph convolutional networks (GCNNs) introduce convolution operations to achieve information transfer and feature transformation between nodes. They aggregate the features of a node's neighbors into the feature representation of the current node, capturing complex relationships between nodes, effectively extracting node features, and enabling deep learning of graph-structured data. The graph convolution operation can be represented as:

[0057] ; in, H (l) It is the first l The node feature matrix of the layer W (l) It is the first l The weight matrix of the layer, , I It is an identity matrix, with the diagonal elements representing the degree of each node. is the activation function. Through the stacking of multiple layers of graph convolutional operations, graph convolutional networks can progressively learn deep feature representations of nodes. Each layer of graph convolution updates the feature representation of the nodes, making it contain richer neighbor information and deeper features.

[0058] Graph Convolutional Autoencoders (GCAEs) are unsupervised learning models that aim to learn low-dimensional vector representations (embeddings) of graphs, typically used for tasks such as node clustering or link prediction. This invention's GCAE consists of an encoder and a decoder. The encoder primarily comprises a Generative Network (GCN). The first-layer GCN aggregates first-order neighbor information for each feature node and performs nonlinear transformations. The second-layer GCN generates low-dimensional vector representations for each feature node, capturing the contextual information of features within the graph structure and fully extracting inter-feature relationships. The decoder reconstructs the original node feature matrix through a two-layer fully connected network, recovering the original feature values ​​from the relationship embeddings. This structure offers the following advantages: ① explicit modeling of inter-feature dependencies; ② unsupervised network feature learning and dimensionality reduction; ③ GCN feature embeddings can learn the global structure of the feature graph, avoiding insufficient local information; ④ the decoder uses Dense layers instead of GCN layers, enabling better global feature reconstruction and capturing complex nonlinear transformations; ⑤ GCAEs and convolutional autoencoders are modally complementary, excelling at capturing global correlation patterns between features; ⑥ analyzing the feature graph and the positions of nodes in the latent space enhances model interpretability.

[0059] (3) Adaptive construction of HI for autoencoders with bidirectional cross-attention mechanism The aforementioned TCAE model excels at capturing local spatial patterns and temporal dependencies within samples; the GCAE branch excels at capturing global correlation patterns between features. To effectively capture the complex interactions between dual-channel information and avoid feature loss and overfitting caused by excessively deep model layers, a cross-attention mechanism adaptive encoder structure is proposed to construct health indicators. A modal information fusion model based on the cross-attention mechanism is established to avoid the problem of simple splicing failing to effectively capture complex interactions between modalities, achieving the fusion of image features and graph structure features and the expression of effective information. An adaptive weight vector is introduced to modulate the features extracted by the model, enhancing features sensitive to degradation, suppressing features sensitive to working conditions and individuals, and other useless features, thereby enhancing the model's adaptability. A cross-attention mechanism autoencoder is established to extract degradation features and automatically construct health indicators (HI).

[0060] Feature fusion via bidirectional cross-attention mechanism: This invention employs Bidirectional cross attention to capture the dependencies between two sequences. The query and key / value pairs come from different sources, and selective information transmission is achieved through attention weights. It can dynamically determine what to focus on based on task requirements, rather than relying on fixed rules. Compared to traditional RNNs or CNNs, it significantly improves efficiency through parallel computation and is suitable for multi-sequence or multi-modal interactive modeling.

[0061] like Figure 10 As shown, this invention establishes a bidirectional cross-attention mechanism to fuse graph convolutional autoencoder features and spatiotemporal convolutional autoencoder features, avoiding the limitations of unidirectional information flow or simple splicing of features. It mainly consists of the following three parts: ① TCAE features search for relevant information from GCAE features, mapping input features to a query (Q), key (K), and value (V) space. Based on attention weights, it determines which parts of the GCAE features should be focused on, and then uses residual connections and layer normalization to prevent gradient vanishing and stabilize the model training process; ② GCAE features search for relevant information from TCAE features, determining which parts of the TCAE features should be focused on based on attention weights, and then uses residual connections and layer normalization to prevent gradient vanishing and stabilize the model training process; ③ The two types of features are connected and expanded to obtain more comprehensive features that interact bidirectionally and mutually enhance each other.

[0062] (4) Adaptive weight vector This invention employs an adaptive weight vector module to extract degradation features, and the weight vector... The calculation formula is: ; in, The Sigmoid activation function can be used to implement a gating mechanism; W 1 and W 2 represents the parameters and weight vectors of the two fully connected (FC) layers. The size is equal to the feature dimension. From Figure 9 As can be seen from this, the weight vector The weight vector after the model is trained is jointly determined by the model parameters and the characteristics of the input signal. Only follow F 1. Change, that is It only changes with the model input signal, therefore, It exhibits better adaptability than fixed weights. Finally, using the formula... Obtain optimized features F 2, F 2 and F 1. They have the same size. Weight vectors Each element in the array is in the range 0-1. This can be viewed as a control gate, which can control the amount of information transmitted by each feature to subsequent layers, enhancing features that are sensitive to degradation trends but insensitive to noise and operating conditions. F 2. It can retain characteristics of good degradation performance.

[0063] (5) Model training To address the problems of existing models struggling to constrain degenerate feature learning from multiple perspectives, leading to low feature quality, poor signal representation, insufficient HI representation ability, and susceptibility to local optima, a multi-constraint TGCAE training loss function is proposed to achieve an organic unity between representation learning and task learning. The total model loss function is:

[0064] ; Loss 1 represents the TCAE reconstruction loss, ensuring that the spatiotemporal convolutional autoencoder can effectively extract spatiotemporal features; where... Loss 2 represents the GCAE reconstruction loss, ensuring its effective learning of graph structure features; Loss 3. To mitigate the loss during fusion and reconstruction, ensure that fusion features can be correctly reconstructed and maintain information integrity; LossThe main objective of this invention is to construct and improve the High-Intensity Array (HI) architecture. To meet the requirements of monotonicity, correlation, robustness, and convexity in rotating machinery degradation, this invention uses a quadratic function to annotate the output features of the encoding network, thereby improving the overall performance of the HI. This invention uses auxiliary tasks 1-3 to enhance the performance of the main task 4, utilizing complementary information between different modal data to improve the model's generalization ability and feature representation quality. To reduce the impact of constraint weights on the model, this invention uses a default weight of 1:1:1:1 for easier model analysis. To enhance the robustness of the loss function to noise, this invention uses the Huber loss function to address the shortcomings of the MAE and MSE loss functions, achieving inverse model adjustment.

[0065] ; in, δ For error hyperparameters, when δ When ~0, Huber's loss will tend towards MAE; when δ As the value approaches infinity (a very large number), the Huber loss tends towards MSE. Therefore, it combines the advantages of both MSE and MAE loss functions.

[0066] Finally, this invention also proposes to eliminate "spurs" in the HI curve using multi-scale Savitzky-Golay filtering. Savitzky-Golay filtering is a filtering method based on local polynomial least squares fitting in the time domain. It can process time-series data without changing the signal trend and width. The multi-scale strategy adopts the principle of smaller window weights (preserving details) and larger window weights (better smoothing effect), which can better balance detail preservation and noise suppression, thereby constructing a smooth and effective HI curve.

[0067] This embodiment verifies and analyzes the effectiveness of the proposed method based on a bearing degradation dataset. The bearing accelerated life test platform consists of an AC motor, a motor speed controller, a shaft, a support bearing, a hydraulic loading system, and a test bearing. Two PCB 352C33 unidirectional accelerometers are fixed to the horizontal and vertical directions of the LDK UER204 rolling bearing, respectively, to collect vibration signals. The sampling frequency is set to 25.6 kHz, the sampling duration is 1.28 s, and the sampling interval between each group of samples is 1 min. Bearing 2-5 is selected for verification under the condition of 2250 rpm rotation speed and 11 kN load.

[0068] This embodiment uses a fusion of correlation, monotonicity, robustness, and stability evaluation methods for feature selection. The results show that: ① Compared with well-performing one-dimensional features, image features also have advantages in degenerate representation ability, and image features account for 48.8% of the total features; ② Among one-dimensional features, RS performs the best, accounting for 26.3% of the total features, while other domain features also have good performance, reflecting the importance of multi-domain features; ③ Among two-dimensional features, time-domain image features perform the best, accounting for 22.5% of the total features, while frequency-domain and time-domain features also perform well, verifying the advantages of multi-domain image features.

[0069] Comparing the results with RMS, Kurtosis, Euclidean, SAE, and DCAE, the overall scores were: RMS 0.5971, Kurtosis 0.5833, Euclidean 0.8107, SAE 0.8453, DCAE 6326, and TGBCAE 0.9207. Compared to the other five methods, TGBCAE improved the scores by 0.3236, 0.3374, 0.1100, 0.0754, and 0.2881, respectively. Compared to RMS, Kurtosis, and Euclidean, the method in this embodiment has more complete feature extraction and stronger nonlinear representation capabilities. Compared to SAE and DCAE, the TGBCAE model improves the model structure by incorporating image features and graph representation structures, and further enhances model performance through adaptive modules and HI constraints.

[0070] The above is one embodiment of the adaptive construction method for rotating machinery health indicators provided in this embodiment. Based on the same idea, this embodiment also provides a corresponding adaptive construction system for rotating machinery health indicators. Specific limitations of the adaptive construction system for rotating machinery health indicators can be found in the limitations of the adaptive construction method for rotating machinery health indicators described above, and will not be repeated here. Each module in the above adaptive construction system for rotating machinery health indicators can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0071] This embodiment also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 An adaptive construction method for rotating machinery health indicators is provided.

[0072] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, database, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0073] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive construction method for health indicators of rotating machinery, characterized in that, The method comprises the following steps: extracting time domain, frequency, time-frequency domain, complexity entropy and relative similarity features of the equipment vibration signal by variational modal decomposition, and converting the equipment vibration signal into time domain, frequency domain and time-frequency domain two-dimensional images to extract gray scale features, gray scale difference statistical features and gray level co-occurrence matrix features, and constructing a feature set; The features are sorted according to the scores by a feature selection method based on correlation, monotonicity, robustness and stability, and the sorted features are displaced row by row and coupled in each row to construct a two-dimensional mixed feature arrangement and coupling sample; According to the feature extraction method, the feature set is divided into multiple types of features, and a feature relationship graph is constructed; wherein the graph nodes are feature types, names and feature values, and the edges are same type information, related information and mutual correlation information; The encoder of the pre-trained spatio-temporal fusion convolutional autoencoder based on Transformer extracts the hidden layer spatio-temporal features of the mixed feature arrangement and coupling sample; the encoder of the pre-trained graph convolutional autoencoder based on GCN extracts the hidden layer graph structure features of the feature relationship graph; The hidden layer spatio-temporal features and the hidden layer graph structure features are fused by the encoder of the bidirectional cross-attention autoencoder to obtain fusion features, the reconstruction features are obtained by the decoder, and the corresponding health indicators are obtained by the quadratic function constraint output.

2. The method of claim 1, wherein the health index of the rotating machinery is adaptively constructed by using the following steps: The encoder part of the spatio-temporal fusion convolutional autoencoder includes a spatial feature extraction module, a Transformer encoder and a deep feature extraction module, and the decoder part adopts an asymmetric decoder including a single deconvolution layer and an up-sampling layer; the spatial feature extraction module includes a two-dimensional convolution layer, a pooling layer and a reshape layer connected in sequence, and adopts a BN layer and a Relu activation function for parameter tuning; the deep feature extraction module includes a convolution layer, a Conv layer and a pooling layer connected in sequence; The spatial feature extraction module extracts the spatial features of the mixed feature arrangement and coupling sample, the Transformer encoder adopts a multi-head attention mechanism to extract global spatio-temporal features from the spatial features, and the deep feature extraction module expands the global spatio-temporal features to obtain the hidden layer spatio-temporal features after spatio-temporal fusion; The encoder part of the graph convolutional autoencoder is two GCN layers, and the decoder part is two fully connected layers; The first layer GCN layer of the encoder aggregates the first-order neighbor information of each feature node of the feature relationship graph and performs nonlinear transformation; the second layer GCN layer generates a low-dimensional vector representation of each feature node of the feature relationship graph to obtain the hidden layer graph structure features; the decoder reconstructs the original node feature matrix from the relationship embedding according to the hidden layer graph structure features to recover the original feature values; The bidirectional cross-attention autoencoder is a double-channel encoder structure based on a bidirectional cross-attention mechanism and connected with an adaptive weight vector module; The double-channel encoder structure maps and fuses the attention mechanisms between the hidden layer spatio-temporal features and the hidden layer graph structure features to obtain fusion features; the adaptive weight vector module dynamically selects the fusion features. According to the spatio-temporal fusion convolutional autoencoder, the graph convolutional autoencoder and the bidirectional cross-attention autoencoder, a health index construction model is constructed.

3. The method of claim 2, wherein the method further comprises: The health index construction model is trained by a training set, and the total loss of the model is: ； In the formula, is a weight, Loss 1 is the reconstruction loss of the spatio-temporal fusion convolutional autoencoder; wherein Loss 2 is the reconstruction loss of the graph convolutional autoencoder; Loss 3 is the feature fusion reconstruction loss in the bidirectional cross-attention autoencoder, which ensures that the fused features can be correctly reconstructed and maintain information integrity; Loss 4 is the health indicator output loss; all use Huber loss function: where delta is an error hyperparameter, and when delta 0, the Huber loss tends to average absolute error; when delta ∞, the Huber loss tends to mean squared error.

4. The adaptive construction method for health indicators of rotating machinery according to claim 1, characterized in that, Further comprising introducing a multi-scale Savitzky-Golay filter to eliminate HI curve burrs and smooth the HI curve.

5. The method of claim 1, wherein the method is characterized by: The time domain features include mean, absolute mean, variance, standard deviation, root amplitude, root mean square, peak value, peak-to-peak value, maximum value and minimum value, and dimensionless features; the dimensionless features include waveform indicators, peak indicators, pulse indicators, margin indicators, skewness indicators and kurtosis indicators; The time-frequency domain features are obtained by 3-layer wavelet packet decomposition of the equipment vibration signal to obtain 8 continuous wavelet energy spectrum features; The complexity entropy features include time domain information entropy, frequency domain information entropy, permutation entropy, sample entropy, approximate entropy, energy entropy and scatter entropy extraction features; The relative similarity features include 5 time domain RS features and 5 frequency domain RS features, which are Euclidean distance, Manhattan distance, Chebyshev distance, cosine distance and Pearson distance.

6. The adaptive construction method for health indicators of rotating machinery according to claim 1, characterized in that, The conversion of the equipment vibration signal into time domain, frequency domain and time-frequency domain two-dimensional images includes the following steps: The sample amplitudes of the time domain vibration signal and the frequency domain vibration signal are normalized to [0, 1] respectively, and then the normalized sample amplitudes are converted into image pixel values to construct time domain two-dimensional images and frequency domain two-dimensional images of the vibration signal; Consider analyzing the vibration signal , is the Dirac function, the vibration signal is transformed from one-dimensional time signal to two-dimensional time-frequency signal by short-time Fourier transform: ； wherein A is the instantaneous amplitude, t is the time, t 0 is a certain time position, is a Gaussian window function, is the frequency; due to the energy distribution of the short-time Fourier transform of the Dirac function is given by: ； The derivative of the frequency of the short-time Fourier transform is calculated: ； For any such that , two-dimensional group delay calculation is: ； The transient extraction operator TEO is introduced: ； Wherein, Delta represents the time support range of the window function, then, the transient extraction transform is represented as: ； The transient extraction transform is represented as: ； Due to Dirac The Dirac delta function has the following properties: ； The transient feature of the extraction transform is converted into a gray scale to obtain a time-frequency domain two-dimensional image. ； The gray difference statistical features include gray mean, gray variance, gray difference statistical mean, gray difference statistical contrast, gray difference statistical image entropy, maximum and minimum values of pixels in the region of interest, pixel value range, pixel value average absolute deviation and pixel value root mean square; 7. The method for adaptive construction of health indices for rotating machinery according to claim 6, characterized in that, The gray level co-occurrence matrix features take the gray level co-occurrence matrixes in 0°, 45°, 90° and 135° directions, and extract the energy, entropy and moment of inertia features of the gray level co-occurrence matrixes in the four directions. The optimal features are selected by the feature selection method based on correlation, monotonicity, robustness and stability indexes, including:

8. The method for adaptive construction of health indices for rotating machinery according to claim 1, characterized in that, Sta ； In the formula, C is the final feature score, is a correlation index, is a monotonicity index, is a robustness index, The system comprises: is a stability index, and the four feature value domains are [0, 1], and the larger the value, the higher the feature score, indicating that the feature can better represent the degradation trend of the device.

9. An adaptive construction system for health indicators of rotating machinery, characterized in that, ​ The feature processing module is configured to extract time domain, frequency, time-frequency domain, complexity entropy and relative similarity features of the equipment vibration signal by variational modal decomposition, and convert the equipment vibration signal into time domain, frequency domain and time-frequency domain two-dimensional images, extract gray scale features, gray scale difference statistical features and gray level co-occurrence matrix features respectively, and construct a feature set; and is further configured to sort the features according to scores by a feature selection method based on correlation, monotonicity, robustness and stability indexes, construct a two-dimensional mixed feature arrangement coupling sample by displacing the sorted features row by row and coupling each row, divide the feature set into multiple types of features according to the feature extraction method, and construct a feature relationship graph; wherein, the nodes of the graph are feature types, names and feature values, and the edges are same type information, related information and mutual correlation information. The health index construction module is configured to extract hidden layer space-time features of the mixed feature arrangement coupling sample by using an encoder of a pre-trained space-time fusion convolutional autoencoder based on a Transformer; extract hidden layer graph structure features of the feature relationship graph by using an encoder of a pre-trained graph convolutional autoencoder based on a GCN; fuse the hidden layer space-time features and the hidden layer space-time features to obtain fusion features by using an encoder of a bidirectional cross-attention autoencoder; reconstruct the fusion features by using a decoder to obtain reconstructed features; and obtain corresponding health indexes by using a quadratic function constraint output.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a data processing program, and the data processing program is executed by the processor to implement the steps of the rotating machinery health index adaptive construction method in any one of claims 1 to 8.

Images