A rotor system fault diagnosis method based on multi-source information chroma space mapping

Abstract

The application provides a rotor system fault diagnosis method based on multi-source information chrominance space mapping, which synchronously collects orthogonal displacement signals and acceleration signals of multiple measuring points of a rotor system, constructs a two-dimensional shaft center track by using the orthogonal displacement signals, maps three acceleration signals into color feature vectors and assigns the shaft center track segments to generate color-coded shaft center tracks of each measuring point; the color-coded shaft center tracks of different measuring points are spliced along the channel dimension to generate a high-dimensional fusion tensor; based on the spatial sparse distribution characteristics of the non-zero color response area in the color-coded shaft center track and the physical meaning of the channel dimension of the high-dimensional fusion tensor, a double-branch convolutional neural network containing a spatial attention convolution branch and a channel attention convolution branch is constructed for feature extraction, and finally, a rotor system fault state recognition result is output. The application realizes the coupled expression of low-frequency displacement track and high-frequency acceleration impact response, fully utilizes the complementarity of multi-source signals, enhances the representation and extraction ability of rotor system fault features, and can accurately recognize the rotor system fault state under complex working conditions.

Description

Technical Field

[0001] This invention belongs to the field of rotating machinery fault diagnosis technology, and specifically relates to a rotor system fault diagnosis method based on multi-source information chromaticity space mapping. Background Technology

[0002] Rotating machinery is a core power component in critical fields such as aerospace, and its operational reliability is crucial for safe production and economic development. Developing fault diagnosis technologies capable of accurately identifying and warning of potential hazards can effectively prevent catastrophic accidents and is of great significance for ensuring the safe and stable operation of equipment in service.

[0003] Most existing intelligent diagnostic methods only use single-sensor signals as input, and noise can overwhelm fault features, leading to a decrease in diagnostic accuracy. However, existing multi-sensor fault diagnosis methods typically only splice together shallow features of one-dimensional vibration signals, ignoring the two-dimensional dynamic response of the rotor system. They fail to perform physical fusion and adaptive feature focusing between heterogeneous signals, making it difficult for the model to fully utilize the complementarity of multi-source information, resulting in insufficient diagnostic robustness and generalization ability under complex working conditions. Summary of the Invention

[0004] To address the shortcomings of existing multi-sensor fault diagnosis methods, such as insufficient utilization of multi-source information and weak robustness, this invention aims to provide a rotor system fault diagnosis method based on multi-source information chromaticity space mapping. Acceleration signals are mapped into color features and embedded into a two-dimensional shaft center trajectory constructed from orthogonal displacement signals. The color-coded shaft center trajectories from different measurement points are then spliced ​​along the channel dimension to generate a color-coded tensor reflecting the physical coupling relationships of multiple sources. Furthermore, a dual-branch convolutional neural network is used to adaptively extract features from the high-dimensional tensor, focusing on key spatial geometric contours and color evolution patterns. This achieves feature enhancement of multi-source heterogeneous signals and suppression of background noise, ultimately enabling high-precision and robust identification of rotor system fault states under complex operating conditions.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A fault diagnosis method for rotor systems based on multi-source information chromaticity space mapping includes the following steps: Step 1: Simultaneously collect multi-source vibration signals of the rotating machinery rotor system at each preset measuring point, including orthogonal displacement signals reflecting the low-frequency characteristics of the equipment, and three acceleration signals reflecting the high-frequency transient impact characteristics, so as to utilize the complementarity of multi-source vibration signals in rotor motion morphology characterization and transient impact response characterization. Step 2: Perform global normalization on the orthogonal displacement signals of each measuring point to maintain the original relative amplitude relationship between the displacement signals of the two orthogonal measurement directions, avoid stretching or compression of the shaft center trajectory caused by separate normalization, preserve the original contour of the shaft center trajectory, and highlight the two-dimensional dynamic response of the rotor system. Step 3: Map the three acceleration signals that reflect the characteristics of high-frequency transient impact to the color space to generate color feature vectors, assign color feature vectors to the line segments of the shaft center trajectory, and perform color aggregation in the overlapping pixel areas to generate the color-coded shaft center trajectory of each measuring point. The pixel information in the color-coded shaft center trajectory is used to simultaneously characterize the spatial position characteristics of the rotor's two-dimensional motion trajectory and the high-frequency acceleration impact response characteristics corresponding to the spatial position characteristics. Step 4: The color-coded shaft center trajectories generated at different measurement points are spliced ​​along the channel dimension to generate a high-dimensional fusion tensor. The different channels of the high-dimensional fusion tensor correspond to the determined measurement points and acceleration signal color responses, so as to retain the difference and complementary information between the responses of different measurement points. The rotor system fault diagnosis database is constructed and divided into training set, validation set and test set. Step 5: Based on the spatial sparse distribution characteristics of non-zero color response regions in the color-coded axis trajectory, and the channel physical meaning of color responses of different measurement points and different acceleration signals corresponding to the channel dimensions of the high-dimensional fusion tensor, a dual-branch convolutional neural network is constructed. The training set data generated in Step 4 is used as input. The spatial attention convolutional branch enhances the feature response of the trajectory line segment region and suppresses the background region by calculating the spatial weight distribution, so as to extract the overall geometric shape of the axis trajectory. At the same time, the channel attention convolutional branch adaptively filters the fault-sensitive responses in different measurement points and different acceleration signal color channels by calculating the channel weights. The dual-branch output features are added element by element and then input into the fully connected neural network to calculate and output the probability of the fault state. Supervised learning is performed using the corresponding fault diagnosis labels to train and update the parameters of the neural network. Step 6: Use the neural network trained in Step 5 as the rotor system fault diagnosis model to process and identify the fault status of the actual monitored rotor system data.

[0006] In step 2, the orthogonal displacement signals of each measuring point are subjected to global normalization processing. Specifically, this process includes: acquiring multi-source vibration signals collected from each preset measuring point, sampling using a sliding window, and for the... m A multi-source state matrix reflecting the dynamic response is constructed from several measurement points. : in, L The length of the sequence truncated in each step of the sliding window. For the first mOrthogonal displacement signals at each measuring point Three acceleration signals were collected synchronously. m The measurement point number; Extract the first one respectively m The global maximum absolute amplitude characteristic value of the orthogonal displacement signal at each measuring point : in, The first m Orthogonal displacement signal at each measuring point The i Each sample value, i For indexing data points, ; use Each data point of the orthogonal displacement signal is scaled proportionally to generate a normalized set of discrete points for the two-dimensional axis-centered trajectory. : in, The first m The first measuring point i Normalized coordinate components of discrete points of the axis trajectory along two orthogonal measurement directions; By using the common maximum absolute amplitude of the two orthogonal displacement signals as a unified scale factor, the normalization process only changes the overall scale of the axis trajectory without changing the relative amplitude relationship between the two orthogonal measurement directions. This avoids stretching or compression of the axis trajectory caused by separate normalization and preserves the original geometric contour of the axis trajectory.

[0007] In step 3, three acceleration signals reflecting high-frequency transient impact characteristics are mapped to a color space to generate color feature vectors. Color feature vectors are assigned to the line segments of the axis-center trajectory, and color aggregation is performed in overlapping pixel areas to generate the color-coded axis-center trajectory for each measurement point. The specific process includes: [The text abruptly ends here, so the translation also ends here.] m Three-channel acceleration sequences at each measurement point Max-min normalization is performed separately to map the values ​​to the range [0, 1], generating the three color components of the corresponding color space. : in, For acceleration sequence number, Acceleration sequence The i Each sample value; By performing maximum-min normalization on the three acceleration sequences respectively, the amplitude scale difference between different acceleration signals is reduced, and they are uniformly mapped to a numerical range suitable for color coding, while preserving the changing trend and relative distribution characteristics of each acceleration sequence within the current sliding window.

[0008] The three generated color components are combined into a color feature vector, then the first color component on the axis-centered trajectory is... i Color feature vector of each point Represented as: Adjacent normalized two-dimensional axisymmetric discrete points and Construct an abstract sequence of trajectory line segments, and calculate the color feature vector of each trajectory line segment. Since each trajectory segment is formed by connecting discrete points of the two-dimensional axis trajectory corresponding to two adjacent sampling times, the color of the trajectory segment is determined by the color feature vectors of the two endpoints. This comprehensively considers the acceleration information corresponding to the two endpoints, thereby more fully characterizing the acceleration response of the rotor when it passes through the spatial position corresponding to the trajectory segment. Convert the color-coded trajectory line segment sequence to a black background and a predefined resolution. The standard three-channel RGB image tensor, this process involves coordinate transformation, rasterization, and composition: Map the endpoints of the trajectory in the normalized continuous space to the discrete pixel coordinate system, and calculate the first... m The first measuring point i Discrete pixel coordinates corresponding to each data point : in, This indicates a round-down operation. Indicates the first m The first measuring point i Normalized coordinate components of discrete points of the axis-centered trajectory along the first orthogonal measurement direction This represents the normalized coordinate components of the discrete points of the axis trajectory along the second orthogonal measurement direction. W and H These represent the number of pixels in the color-coded axisymmetric trajectory image along the horizontal and vertical directions, respectively. ; Linear interpolation is used to rasterize the mapped trajectory segments; for spatially overlapping regions, the coordinates of the segments passing through the same discrete pixel are rasterized. The colors of all trajectory line segments are aggregated to determine the final composite color of the pixel coordinate. : in, Represents discrete pixel coordinates The set of indices of all trajectory segments intersecting spatial regions. The total number of line segments in the index set. Indicates the first index in the set k The color feature vector of intersecting line segments.

[0009] In step 4, the color-coded axis trajectories generated from different measuring points are stitched together along the channel dimension. This preserves the spatial structure of the color-coded axis trajectories at each measuring point while retaining the differences in vibration response collected at different spatial locations. By utilizing the complementarity between the vibration information from multiple measuring points, a high-dimensional fusion tensor is generated. Construct a rotor system fault diagnosis database: in, Indicates the first m The color-coded axisymmetric trajectory tensor generated at each measurement point; All high-dimensional fusion tensors extracted under different rotor fault states Assign corresponding fault status labels to construct a rotor system fault diagnosis database.

[0010] In step 5, the color-coded axis trajectory includes a non-zero color response region formed by trajectory segments and a background region not traversed by the trajectory segments. The non-zero color response region is sparsely distributed in the image space. The high-dimensional fusion tensor channel dimension is formed by splicing the color-coded axis trajectories of different measurement points, so that each channel corresponds to the color response of different measurement points and different acceleration signals. Based on the spatial sparse distribution characteristics of the color-coded axis trajectory and the channel physical meaning of the high-dimensional fusion tensor, the high-dimensional fusion tensor is... Parallel input is fed to spatial attention convolution branches and channel attention convolution branches for feature extraction and rotor system fault diagnosis: In the spatial attention convolution branch, the spatial weight matrix of the tensor is first calculated using the spatial attention module. Extract the overall contour features of the axis trajectory of multiple measurement points: in, It is a non-linear activation function. The kernel size is Convolution operations; Since the effective vibration response in the color-coded axis trajectory is mainly concentrated in the non-zero color response area occupied by the trajectory segment, while the background area is a zero-value area, the spatial attention module is used to calculate the weight of each spatial position, which can enhance the feature response of the area where the trajectory segment is located and suppress the interference of the background area, thus avoiding a large number of invalid background pixels from weakening the geometric contour and local morphological features of the axis trajectory.

[0011] In the channel attention convolution branch, the channel weight matrix of the feature channel dimension is calculated. Extract the relative response weights of different acceleration signals corresponding to each color channel: in, This is the weight matrix for the first layer of the multilayer perceptron, used to weight the pooled components. C 3D channel feature mapping to C / r dimension, This is the weight matrix for the second layer of the multilayer perceptron, used to... C / r 3D channel features restored to C dimension, C for Number of channels r This is the channel dimensionality reduction ratio. and These represent max pooling and average pooling operations, respectively. Since the different channels of the high-dimensional fusion tensor correspond to the determined measurement points and acceleration signal color responses, and the response levels of different measurement points and different acceleration signals to different fault states vary, the channel attention module is used to calculate the relative response weights of each channel. This can enhance the fault-sensitive channels related to the current fault state and suppress redundant channels or channels affected by noise, thus avoiding the weakening of key fault responses caused by equal weighting of each channel.

[0012] The calculated channel weight matrix and spatial weight matrix are respectively fused with the high-dimensional fusion tensor of the input features. Element-wise weighted multiplication is performed, and then each branch is independently fed into its respective convolutional neural network to extract deep features. The output features of the two branches are then summed element-wise to obtain a high-dimensional joint feature representation that combines global geometric contours and local color changes. H : in, This represents element-wise multiplication of tensors. and These represent the multi-layer convolutional neural network mapping functions corresponding to the spatial branch and the multi-layer convolutional neural network mapping functions corresponding to the channel branch, respectively. High-dimensional joint feature representation H After being flattened, the input is fed into a fully connected classification neural network, which maps the output to the probability distribution of the preset rotor fault categories, and the result with the highest probability is used as the final fault diagnosis label of the rotor system.

[0013] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention utilizes orthogonal displacement signals to construct the rotor shaft center trajectory, highlighting the two-dimensional dynamic response of the system. Through chromaticity space mapping, the synchronously acquired acceleration signals are mapped into color feature vectors and embedded into the shaft center trajectory segments. This allows the color-coded shaft center trajectory to simultaneously contain the spatial motion position features corresponding to low-frequency displacement signals and the impact response features corresponding to high-frequency acceleration signals. Compared to diagnostic methods that only use ordinary shaft center trajectories or one-dimensional vibration signal splicing, this invention achieves the physical coupling expression of multi-source heterogeneous vibration signals in the same image space. It can simultaneously characterize motion morphology changes caused by low-frequency faults such as imbalance, misalignment, and loosening, as well as high-frequency acceleration response changes caused by impact faults such as rubbing and bearing damage.

[0014] (2) This invention generates a high-dimensional fusion tensor by splicing the color-coded shaft center trajectory generated by different measuring points along the channel dimension, making full use of the response differences and complementary information between different measuring points, adapting to the fault diagnosis needs of rotor systems under different numbers of measuring points and complex industrial monitoring scenarios, and improving the engineering applicability and scalability of the method.

[0015] (3) This invention addresses the characteristics of the effective trajectory response region in the color-coded shaft center trajectory being spatially sparsely distributed and the channel dimensions having clear physical meanings. It constructs spatial attention convolution branches and channel attention convolution branches for feature extraction. The spatial attention convolution branch is used to capture the trajectory contour, while the channel attention convolution branch is used to adaptively determine the contribution weights of the color response channels corresponding to different measurement points and different acceleration signals. This enhances the feature extraction capability of the effective response region of the shaft center trajectory and the fault-sensitive color channels, thereby improving the identification capability of rotor fault states under complex working conditions. Attached Figure Description

[0016] Figure 1 This is a flowchart of a rotor system fault diagnosis method based on multi-source information chromaticity space mapping.

[0017] Figure 2 A schematic diagram for generating the color-coded axis trajectory of a single measurement point.

[0018] Figure 3 This is the confusion matrix for the test set. Detailed Implementation

[0019] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings and examples.

[0020] This invention proposes a rotor system fault diagnosis method based on multi-source information chromaticity space mapping. First, orthogonal displacement and acceleration signals from multiple measuring points of the rotor system are simultaneously acquired. Second, the displacement signals reflecting low-frequency characteristics are constructed into a two-dimensional shaft center trajectory, and the acceleration signals reflecting high-frequency impact characteristics are mapped as color changes and embedded within it, generating color-coded shaft center trajectories for each measuring point. Subsequently, the trajectories generated from different measuring points are concatenated along the channel dimension to generate a high-dimensional fusion tensor. Finally, the high-dimensional fusion tensor is input into a two-branch convolutional neural network containing parallel channel attention convolution branches and spatial attention convolution branches, adaptively extracting the spatial contour features and the relative response weights of different acceleration signals corresponding to each color channel, achieving accurate identification of rotor system fault states. Experimental results show that the method of this invention can accurately identify rubbing faults at different locations, providing important technical support for the intelligent operation and maintenance of rotating machinery.

[0021] Taking the identification of rubbing position on a sliding bearing rotor test bench as an example, the specific training and testing process of this rubbing positioning method is as follows: Figure 1 As shown, it includes the following steps: Step 1: Set up a measuring point at each end of the test bench. At each measuring point, a triaxial acceleration sensor is placed on the bearing housing, and an orthogonal eddy current displacement sensor is placed inside the bearing housing. Simultaneously collect multi-source vibration signals under different contact conditions and without contact conditions.

[0022] Step 2: Use a sliding window sampling method to sample the multi-source vibration signals synchronously acquired in Step 1. The sample is divided to obtain multi-source vibration signal samples corresponding to each measuring point. For example... Figure 2 As shown, for each measuring point, the global maximum absolute amplitude feature value of the orthogonal displacement signal at each measuring point is extracted. The displacement is scaled proportionally using this eigenvalue to generate a normalized two-dimensional axisymmetric trajectory discrete point set that retains the original contour.

[0023] in, This is the measurement point number.

[0024] Step 3: According to Figure 2 The single-point color-coded axis trajectory construction process shown in the diagram performs maximum-minimum normalization on the different acceleration signals at each measuring point and maps them to... Within the range, it is used as the RGB color component. The color of the trajectory line segment formed by adjacent trajectory data points is calculated. It is derived from the average color of the two endpoints: Set the standard image resolution to Map the trajectory line segment sequence to a discrete pixel coordinate system, the first... m The first measuring point i Discrete pixel coordinates corresponding to each data point The calculation is as follows: Linear interpolation is used to rasterize the mapped trajectory segments, and spatially overlapping regions that pass through the same discrete pixel are processed. Aggregate the colors of all trajectory line segments to generate the dimensions of each measuring point. Standard color-coded axisymmetric trajectory tensor: Step 4: Concatenate the color-coded axisymmetric trajectory tensors generated from the left and right end measurement points along the channel dimension to generate a tensor of size [size missing]. High-dimensional fusion tensor : All high-dimensional fusion tensors extracted from 12 different rubbing locations and 1 no-rubbing state are assigned fault labels, and a database is constructed and divided into training set, validation set and test set.

[0025] Step 5: Construct a two-branch convolutional neural network containing spatial attention convolutional branches and channel attention convolutional branches. (The generated...) High-dimensional fusion tensor The initial input features are fed into both branches in parallel.

[0026] In the spatial attention convolution branch, the spatial attention module is first used to calculate the spatial weight distribution of the tensor. : The weight matrix of the feature channel dimension is calculated in the channel attention convolution branch. Extract the relative response weights of different acceleration signals corresponding to each color channel. Number of channels Channel dimensionality reduction ratio : The above weight matrix is ​​fused with the high-dimensional input feature tensor. After element-wise weighted multiplication, the features are fed into their respective multi-layer convolutional neural networks to extract deep features. In this embodiment, the convolutional neural network of the independent branch consists of several cascaded convolutional modules, each of which sequentially includes a two-dimensional convolutional layer, a non-linear activation layer, and a pooling layer. The deep feature maps output by the two branches are element-wise added and fused into high-dimensional joint features, which are then flattened and input into a fully connected neural network layer. Finally, the probability distribution of 13 running states (corresponding to 12 collision locations and 1 no-collision state) is mapped and output through the Softmax function, and the network training is completed using the cross-entropy loss function.

[0027] Step 6: Use the neural network trained in Step 5 as the rubbing location model to perform rubbing location of the rotor system. In this embodiment, the test set includes 12 different rubbing position states and no-rubbing state of the sliding bearing rotor system, totaling 13 operating states. Each state contains 30 test samples, for a total of 390 test samples. Of these, 388 samples were correctly identified, and the test set identification accuracy rate was 99.49%.

[0028] To further illustrate the recognition effect of the method of the present invention on different collision position states, the classification results of the test set were statistically analyzed, and the confusion matrix is ​​as follows: Figure 3 As shown, the vertical axis represents the true label, the horizontal axis represents the predicted label, and the diagonal elements represent the number of samples correctly identified for each state. Most samples were accurately classified, with only a small number of misclassifications between a few closely spaced rubbing positions. This indicates that the method of this invention has high distinguishing ability and identification accuracy for faults at different rubbing positions, and can provide a basis for monitoring the operating status and making maintenance decisions for rotor systems.

Claims

1. A fault diagnosis method for rotor systems based on multi-source information chromaticity space mapping, characterized in that, Includes the following steps: Step 1: Simultaneously collect multi-source vibration signals of the rotating machinery rotor system at each preset measuring point, including orthogonal displacement signals reflecting the low-frequency characteristics of the equipment, and three acceleration signals reflecting the high-frequency transient impact characteristics. Step 2: Perform global normalization on the orthogonal displacement signals of each measuring point to preserve the original contour of the shaft center trajectory and highlight the two-dimensional dynamic response of the rotor system; Step 3: Map the three acceleration signals that reflect the characteristics of high-frequency transient impact to the color space to generate color feature vectors, assign color feature vectors to the line segments of the shaft center trajectory, and perform color aggregation in the overlapping pixel areas to generate the color-coded shaft center trajectory of each measuring point. The pixel information in the color-coded shaft center trajectory is used to simultaneously characterize the spatial position characteristics of the rotor's two-dimensional motion trajectory and the high-frequency acceleration impact response characteristics corresponding to the spatial position characteristics. Step 4: The color-coded shaft center trajectories generated from different measurement points are spliced ​​along the channel dimension to generate a high-dimensional fusion tensor, constructing a rotor system fault diagnosis database, and dividing the rotor system fault diagnosis database into a training set, a validation set, and a test set; Step 5: Based on the spatial sparse distribution characteristics of non-zero color response regions in the color-coded axis trajectory, and the channel physical meaning of color responses of different measurement points and different acceleration signals corresponding to the channel dimensions of the high-dimensional fusion tensor, a dual-branch convolutional neural network is constructed. The training set data generated in Step 4 is used as input. The spatial attention convolutional branch adaptively extracts contour features by calculating the spatial weight distribution, while the channel attention convolutional branch adaptively extracts fault-sensitive responses of color channels corresponding to different measurement points and different acceleration signals by calculating the channel weights. The output features of the two branches are added element by element and then input into the fully connected neural network to calculate and output the probability of the fault state. Supervised learning is performed using the corresponding fault diagnosis labels to train and update the parameters of the neural network. Step 6: Use the neural network trained in Step 5 as the rotor system fault diagnosis model to process and identify the fault status of the actual monitored rotor system data.

2. The rotor system fault diagnosis method based on multi-source information chromaticity space mapping according to claim 1, characterized in that, In step 2, the orthogonal displacement signals of each measuring point are subjected to global normalization processing. Specifically, this process includes: acquiring multi-source vibration signals collected from each preset measuring point, sampling using a sliding window, and for the... m A multi-source state matrix reflecting the dynamic response is constructed from several measurement points. : in, L The length of the sequence truncated in each step of the sliding window. For the first m Orthogonal displacement signals at each measuring point Three acceleration signals were collected synchronously. m The measurement point number; Extract the first one respectively m The global maximum absolute amplitude characteristic value of the orthogonal displacement signal at each measuring point : in, The first m Orthogonal displacement signal at each measuring point The i Each sample value, i For indexing data points, ; use Each data point of the orthogonal displacement signal is scaled proportionally to generate a normalized set of discrete points for the two-dimensional axis-centered trajectory. : in, The first m The first measuring point i Normalized coordinate components of discrete points of the axis trajectory along two orthogonal measurement directions.

3. The rotor system fault diagnosis method based on multi-source information chromaticity space mapping according to claim 1, characterized in that, In step 3, three acceleration signals reflecting high-frequency transient impact characteristics are mapped to a color space to generate color feature vectors. Color feature vectors are assigned to the line segments of the axis-center trajectory, and color aggregation is performed in overlapping pixel areas to generate the color-coded axis-center trajectory for each measurement point. The specific process includes: [The text abruptly ends here, so the translation also ends here.] m Three-channel acceleration sequences at each measurement point Max-min normalization is performed separately to map the values ​​to the range [0, 1], generating the three color components of the corresponding color space. : in, For acceleration sequence number, Acceleration sequence The i Each sample value; The three generated color components are combined into a color feature vector, then the first color component on the axis-centered trajectory is... i Color feature vector of each point Represented as: Adjacent normalized two-dimensional axisymmetric discrete points and Construct an abstract sequence of trajectory line segments, and calculate the color feature vector of each trajectory line segment. The color feature vector of each line segment is obtained by calculating the average of the color feature vectors of its two endpoints: Convert the color-coded trajectory line segment sequence to a black background and a predefined resolution. The standard three-channel RGB image tensor, this process involves coordinate transformation, rasterization, and composition: Map the endpoints of the trajectory in the normalized continuous space to the discrete pixel coordinate system, and calculate the first... m The first measuring point i Discrete pixel coordinates corresponding to each data point : in, This indicates a round-down operation. Indicates the first m The first measuring point i Normalized coordinate components of discrete points of the axis-centered trajectory along the first orthogonal measurement direction This represents the normalized coordinate components of the discrete points of the axis trajectory along the second orthogonal measurement direction. W and H These represent the number of pixels in the color-coded axisymmetric trajectory image along the horizontal and vertical directions, respectively. ; Linear interpolation is used to rasterize the mapped trajectory segments; for spatially overlapping regions, the coordinates of the segments passing through the same discrete pixel are rasterized. The colors of all trajectory line segments are aggregated to determine the final composite color of the pixel coordinate. : in, Represents discrete pixel coordinates The set of indices of all trajectory segments intersecting spatial regions. The total number of line segments in the index set. Indicates the index set of the first k The feature vector of line segment color of intersecting line segments.

4. The rotor system fault diagnosis method based on multi-source information chromaticity space mapping according to claim 1, characterized in that, In step 4, the color-coded axis trajectories generated from different measurement points are spliced ​​along the channel dimension to generate a high-dimensional fusion tensor. Construct a rotor system fault diagnosis database: in, Indicates the first m The color-coded axisymmetric trajectory tensor generated at each measurement point; All high-dimensional fusion tensors extracted under different rotor fault states Assign corresponding fault status labels to construct a rotor system fault diagnosis database.

5. The rotor system fault diagnosis method based on multi-source information chromaticity space mapping according to claim 1, characterized in that, In step 5, the color-coded axis trajectory includes a non-zero color response region formed by trajectory segments and a background region not traversed by the trajectory segments. The non-zero color response region is sparsely distributed in the image space. The channel dimension of the high-dimensional fusion tensor is formed by splicing the color-coded axis trajectories of different measurement points along the channel dimension, so that each channel corresponds to the color response of different measurement points and different acceleration signals. Based on the spatial sparse distribution characteristics of the color-coded axis trajectory and the channel physical meaning of the high-dimensional fusion tensor, the high-dimensional fusion tensor is... Parallel input is fed to spatial attention convolution branches and channel attention convolution branches for feature extraction and rotor system fault diagnosis: In the spatial attention convolution branch, the spatial weight matrix of the tensor is first calculated using the spatial attention module. Extract the overall contour features of the axis trajectory of multiple measurement points: in, It is a non-linear activation function. The kernel size is Convolution operations; In the channel attention convolution branch, the channel weight matrix of the feature channel dimension is calculated. Extract the relative response weights of different acceleration signals corresponding to each color channel: in, This is the weight matrix for the first layer of the multilayer perceptron, used to weight the pooled components. C 3D channel feature mapping to C / r dimension, This is the weight matrix for the second layer of the multilayer perceptron, used to... C / r 3D channel features restored to C dimension, C for Number of channels r This is the channel dimensionality reduction ratio. and These represent max pooling and average pooling operations, respectively. The calculated channel weight matrix and spatial weight matrix are respectively fused with the high-dimensional fusion tensor of the input features. Element-wise weighted multiplication is performed, and then each branch is independently fed into its respective convolutional neural network to extract deep features. The output features of the two branches are then summed element-wise to obtain a high-dimensional joint feature representation that combines global geometric contours and local color changes. H : in, This represents element-wise multiplication of tensors. and These represent the multi-layer convolutional neural network mapping functions corresponding to the spatial branch and the multi-layer convolutional neural network mapping functions corresponding to the channel branch, respectively. High-dimensional joint feature representation H After being flattened, the input is fed into a fully connected classification neural network, which maps the output to the probability distribution of the preset rotor fault categories, and the result with the highest probability is used as the final fault diagnosis label of the rotor system.

Images