Phase modulator rotor fault diagnosis system based on multi-receptive field information fusion cnn
By fusing multi-sensor vibration data and using the SDP-MRFRCNN network structure, the problem of insufficient multi-scale feature extraction and fusion in the vibration fault diagnosis of condenser cameras is solved, and efficient and accurate fault identification and classification are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-15
- Publication Date
- 2026-03-24
Smart Images
Figure CN116561615B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of phase modifier fault diagnosis, and in particular to a multi-receptive field information fusion convolutional neural network phase modifier rotor fault diagnosis method and an acceleration signal measuring point position arrangement method based on SDP-MRFRCNN. BACKGROUND
[0002] Nowadays, new energy such as wind power and photovoltaic is being massively connected to the grid, and ultra-high voltage direct current transmission technology is also developing, and the demand for reactive power regulation of the power grid is gradually increasing. As a large-scale reactive power regulation device, the reactive power increase and absorption capacity of the phase modifier are strong, which can enhance the dynamic voltage regulation capability of the power grid. With excellent dynamic reactive power compensation capability, the phase modifier can eliminate the adverse effects of sudden voltage increase in the power system, especially at the grid side, the phase modifier can quickly and massively absorb the reactive power generated by commutation failure, and quickly compensate for the reactive power to ensure the stability of the system reactive power. Therefore, the safe and stable operation of the phase modifier is particularly critical.
[0003] As a typical rotating mechanical equipment, the rotor is the most important component of the phase modifier, and if a fault occurs, it will cause the unit to shut down, resulting in significant economic losses, and even threatening personal safety. Therefore, accurate and rapid intelligent fault diagnosis of the phase modifier rotor is of great significance to improve the safety and economy of the unit. For large rotating machinery, vibration sensors are arranged at multiple measuring points, which are economical and accurate, so analyzing vibration signals is a common method of condition monitoring. In order to overcome external interference and improve the diagnostic performance of the model under the influence of noise and non-stationary conditions, the optimization of the current fault diagnosis method based on traditional deep learning mainly deepens the learning depth by complicating the model structure. However, the convolution and pooling layers of the convolutional neural network are stacked, which lacks consideration of extracting fault features of different scales and directions, and the overcomplicated model stacking will reduce the training speed and consume a lot of computing power. From the perspective of improving training speed, classification accuracy and model adaptability to multi-scale characteristics of vibration signals, how to combine data visualization, multi-scale feature extraction and model feature learning ability enhancement is the starting point of the present application. SUMMARY
[0004] Therefore, the present application provides a phase modifier rotor fault diagnosis method and an acceleration signal measuring point position arrangement method to solve the problems of poor multi-scale feature extraction fusion, unsatisfactory classification effect and deep learning network structure optimization in the prior art.
[0005] To achieve the above object, the present application adopts the following technical solutions:
[0006] In a first aspect, the present application provides a multi-sensor vibration data fusion method, which comprises:
[0007] For the multi-dimensional distribution of the phase modifier vibration data, the radial horizontal and axial vertical directions of the excitation and disc turning end of the phase modifier are set as the arrangement direction of the sensor measuring points, and multi-channel synchronous monitoring technology is used to perform multi-sensor fusion sampling on the vibration signals under different working conditions to obtain multi-dimensional high-frequency sampling data with a single learning sample length of 20480*4.
[0008] When the existing method performs data sampling on the vibration time domain signal, a single channel sampling method is often used, which may cause the singleness of the data mode and sometimes cause the loss of key fault information. In addition, the small number of channels may also cause the lack of data set, and the single data structure may even cause the deep learning model to fall into the risk of overfitting. In view of this, the present application designs a multi-sensor vibration data fusion method, which performs multi-sensor fusion sampling at a specific position and direction through multiple sensors, so as to achieve multi-scale data fusion sampling and sufficiently suppress the risk of overfitting in the deep learning convolutional neural network learning.
[0009] In a second aspect, the embodiments of the present application provide a multi-scale visualized fusion feature method, which comprises:
[0010] The multi-dimensional fusion vibration signals collected by the multiple sensors are converted into scatter points in polar coordinates through the SDP method, and then the four sensor signals are taken as inputs to form a pair of petals in each quadrant of the polar coordinates. Finally, the vibration information of each sensor is visualized and fused. The scatter plots of the multiple sensors are fused to generate a single SDP image containing multi-source information, which highlights the differences between different signals.
[0011] (1)
[0012] (2)
[0013] (3)
[0014] In the formula, the maximum value and the minimum value of the waveform amplitude of the original signal correspond to and respectively. is the polar coordinate radius of the point; and are the angles of the point rotating in the counterclockwise and clockwise directions about the mirror symmetry plane respectively; is the specified mirror symmetry plane rotation angle , m = 1, 2, …, , is the number of mirror symmetry planes; is the gain coefficient ).
[0015] Currently, there are few visualized information fusion methods in the field of fault diagnosis. The method simultaneously converts the original signals of multiple sensors through the SDP imaging method, realizes sufficient fusion of multi-source information, and makes a single SDP picture simultaneously include the original vibration information and feature extraction information of multiple sensors, thereby containing rich device operation state features and highlighting the subtle differences between different signal faults.
[0016] In a third aspect, the application provides a SDP-MRFRCNN-based multi-receptive field information fusion convolutional neural network rotor fault diagnosis network structure for a phase modifier, and the content includes:
[0017] First, a CBS module is built. A batch normalization (BN) layer is added after the convolution operation to standardize the samples, and a silu activation function is used to replace the convolution operation in the traditional convolutional neural network, so as to obtain a matrix of effective activation weights. Secondly, a fast spatial pyramid pooling SPPF module is constructed. The matrix extracted by the CBS module is subjected to 1×1, 3×3 and 5×5 subgraph division pooling operations respectively, and the feature matrices of the three scales are connected to fuse the feature matrices of the three different scales. Finally, in view of the shortcomings existing in the structure of the traditional CNN and in combination with the complex features existing in the fault information of the phase modifier, a mixed receptive field residual convolutional neural network structure MRFRCNN is proposed.
[0018] For each fault condition, the vibration data of the measurement position points under the current condition are subjected to multi-sensor visualized feature fusion data set as training samples; the proposed SDP-MRFRCNN is applied to train the training samples of each region, and the corresponding fault categories of each region are obtained respectively;
[0019] In a fourth aspect, the application provides a clustering analysis method for evaluating the classification effect of the rotor fault diagnosis deep learning network of the phase modifier, and the content includes: adopting a T-SNE method to perform dimensionality reduction processing on the full connection output layer data of the classified test set data, to generate a clustering cluster, so that the result is further visualized, as shown in Figure 6 It can be seen that the classification effect of the SDP-MRFRCNN network structure for the three operating states of the phase modifier is very significant, the feature boundaries between different categories are obvious, the feature distribution between the same categories is compact, and the classification accuracy of the deep learning network proposed in the patent is effectively visualized and evaluated.
[0020] The application also provides a SDP-MRFRCNN-based multi-receptive field information fusion convolutional neural network rotor fault diagnosis method for a phase modifier, and the specific steps include:
[0021] S1: Obtain rotor multi-sensor vibration data of different fault types of the phase modifier device under variable working condition, and the sensor test points are: phase modifier rotor disc end X, phase modifier rotor disc end Y, phase modifier rotor excitation end X, and phase modifier rotor excitation end Y, which are sequentially numbered as No. 1 sensor, No. 2 sensor, No. 3 sensor, and No. 4 sensor. The vibration signals of the equipment under normal, rubbing, and unbalance conditions are collected.
[0022] S2: Perform information fusion on the vibration signals of multiple sensors of the phase modifier through the symmetric point mode. First, determine the polar coordinate radius, then determine the polar coordinate rotation angle counterclockwise along the initial line, and finally determine the polar coordinate rotation angle clockwise along the initial line. In the process of multi-sensor information fusion, the original signals of each sensor of the phase modifier in three states are decomposed by SDP and displayed in the first quadrant. Then, the vibration characteristics of the four sensors are fused, and a pair of petals is formed in each of the four quadrants of the polar coordinate.
[0023] S3: Obtain the image of fused multi-source vibration information.
[0024] The multi-dimensional fused vibration signals collected by multiple sensors are converted into scattered points in polar coordinates by the SDP method. Then, the signals of No. 1, No. 2, No. 3, and No. 4 sensors are taken as inputs, and mirror symmetry 45º, 135º, 225º, and 315º are set with a parameter of 45°. A pair of petals is formed in each of the four quadrants of the polar coordinate. Finally, the vibration information of each sensor is fused, the scatter plots of multiple sensors are fused, and a single SDP image containing multi-source information is generated.
[0025] S4: Build a hybrid receptive field residual convolutional neural network model to learn the fault features of the phase modifier rotor by information fusion and multi-receptive field residual network structure.
[0026] The steps of building a hybrid receptive field residual convolutional neural network model are as follows: First, build a CBS module. After convolution operation, add a batch normalization (BN) layer to standardize the samples, and use a silu activation function to replace the convolution operation in the traditional convolutional neural network to obtain a matrix of effective activation weights. Second, build a fast spatial pyramid pooling SPPF module. Perform 1x1, 3x3, and 5x5 subgraph division pooling operations on the matrix extracted by the CBS module, and connect the feature matrices of the three scales to fuse the feature matrices of the three different scales. Finally, in view of the shortcomings of the traditional CNN structure and the complex features of the phase modifier fault information, a hybrid receptive field residual convolutional neural network structure MRFRCNN is proposed.
[0027] S5: Construct a multi-sensor fusion atlas dataset sample.
[0028] Create an image dataset for model training and testing, and generate 250 pieces of data for each of the normal, rubbing and imbalance states through information fusion of the modulator vibration SDP image.
[0029] S6: Data division and model training;
[0030] After mixing the image dataset as a whole, set the random number to shuffle the data, divide the training set and the test set in a ratio of 4:1, and sequentially encode the three states as 0, 1 and 2 before inputting the network; set the training configuration, set the network learning rate to 0.0001, the batch sample number to 32, the iteration to 50 rounds, the length of the input modulator one-dimensional vibration sequence to Lin, and the size of the SDP multi-sensor feature fusion two-dimensional image matrix to a*b in the Pytorch environment.
[0031] S7: Modulator rotor fault diagnosis, input the vibration sequence to be tested, and output the matrix sequence label to be classified.
[0032] The input layer of the mixed receptive field residual convolutional neural network structure MRFRCNN is the original image of 128x128, after passing through the CBS layer of the first 3x3 convolution kernel, the data is divided into three branches, branch 1: 1x1 convolution CBS branch; branch 2: 3x3 CBS branch; branch 3: 5x5 CBS branch; all three branches use two CBS layers, the kernel number is 16 and 32, and the step is 2; the outputs of branch 2 and branch 3 are subjected to ADD operation, then pass through the CBS layer of 1x1 point convolution to make the data flow smooth, and then perform CONCAT operation with the output of branch 1, that is, form a residual connection with branch 1, reduce the model complexity to prevent overfitting, and prevent gradient disappearance or gradient explosion; after the CONCAT of the three branches, the SPPF module is used for dimension reduction, multiple windows (pooling window) are used to extract spatial feature information of different sizes, and the robustness of the model to spatial layout and object deformation is improved; finally, two fully connected layers are used, and Softmax is used to realize multi-classification.
[0033] This invention adopts the above technical solutions, comprehensively considering the holistic nature of multi-sensor information fusion for extracting fault features of the camera condenser, as well as the visualization and fusion of feature information, and the depth and accuracy of deep learning networks. It proposes a fault diagnosis method based on SDP-MRFCNN, achieving efficient, accurate, and automatic identification of the camera condenser rotor state. The model boasts high accuracy, fewer iterations, and fast convergence, significantly improving the learning effect and state identification accuracy of camera condenser vibration signal features. The SDP method fuses information from multiple sensors, mapping one-dimensional signals into a visualized two-dimensional scatter image, avoiding feature loss caused by a single information source. A single SDP image contains vibration information from four sensors in different orientations, more clearly, intuitively, and holistically reflecting the vibration characteristics of the equipment, greatly enhancing the feature differences between different operating states. The MRFCNN network abandons the traditional CNN network structure of stacking numerous single-branch convolutional and pooling layers, employing multi-receptive-field CBS layers, SPPF pooling, and residual blocks, enabling the joint extraction of camera condenser vibration features from different receptive fields. It solves the problem of extracting repetitive features related to graphs in traditional CNNs, greatly improving the speed and accuracy of classification, saving computational costs, and avoiding overfitting. Attached Figure Description
[0034] Figure 1 This is a sensor measurement point arrangement scheme for a camera condenser provided in an embodiment of the present invention;
[0035] Figure 2 This is a schematic diagram of a symmetric point mode SDP applicable in an embodiment of the present invention;
[0036] Figure 3 This is a deep learning CBS layer structure diagram applicable to an embodiment of the present invention;
[0037] Figure 4 This is a schematic diagram of an SPPF module structure applicable to an embodiment of the present invention;
[0038] Figure 5 This is a schematic diagram of an MRFRCNN network architecture provided in an embodiment of this application;
[0039] Figure 6 This is a T-SNE diagram for classifying fault samples of a synchronous condenser provided in an embodiment of this application;
[0040] Figure 7 This is a flowchart of a synchronous condenser rotor fault diagnosis system provided in an embodiment of this application. Detailed Implementation
[0041] (1) In the aspect of data enhancement, step 1 is based on the original vibration data of the phase modifier under different working conditions. Through multi-sensor information fusion, 4-dimensional synchronous fusion acquisition is used to mix the data set, avoiding the single data mode caused by single sensor acquisition method. Through multi-scale fusion to expand the data set, rich original data is provided for the subsequent steps, and the risk of overfitting caused by learning single data source by CNN network is fully suppressed.
[0042] (2) In the aspect of fault feature extraction, step 2 uses the SDP method to fuse the vibration signals of multiple sensors of the phase modifier through the symmetry point mode. First, the polar coordinate radius is determined, then the polar coordinate rotation angle along the initial line is determined, and finally the polar coordinate rotation angle along the initial line is determined. The 1st, 2nd, 3rd and 4th sensors are used as input, respectively, to form a pair of petals in each quadrant of the polar coordinate, and finally the vibration information of each sensor is fused to obtain the image of fused multi-source vibration information. The completeness and distinguishability of the overall prominent state feature are improved, thus making up for the information loss caused by single feature component.
[0043] (3) In the aspect of feature recognition, steps 3-4 use the SDP imaging method to effectively fuse the original signals of multiple sensors, realizing the complementation of multi-source information. Then the SDP image of fused multi-source features is input into the MRFRCNN model, and through the powerful feature dimension reduction extraction learning of the network structure, fast and automatic recognition is achieved. Avoiding the subjective interference caused by manual recognition, it truly realizes the intelligence and can effectively diagnose the on-site fault, truly realizes the efficiency and reduces the cost.
[0044] (4) In the aspect of verifying the classification accuracy and effect of the deep learning network structure, step 5 uses the T-SNE method to reduce the dimension of the full connection output layer data after classification of the test set data, generates clustering clusters, and makes the results further visualized.
Claims
1. A method for diagnosing camera rotor faults based on a multi-receptive-field information fusion convolutional neural network using SDP-MRFRCNN, characterized in that, Includes the following steps: S1: Obtain multi-sensor vibration data of rotors under different fault types under varying operating conditions of the synchronous condenser equipment. The sensor test points are: synchronous condenser rotor turning end X, synchronous condenser rotor turning end Y, synchronous condenser rotor excitation end X, and synchronous condenser rotor excitation end Y, which are numbered as sensor 1, sensor 2, sensor 3, and sensor 4 in sequence. S2: By fusing the vibration signals from multiple sensors of the camera condenser using the symmetrical point mode, first determine the polar coordinate radius, then determine the counterclockwise rotation angle of the polar coordinate along the initial line, and finally determine the clockwise rotation angle of the polar coordinate along the initial line. S3: Acquire an image that fuses multi-source vibration information. Using the signals from sensors 1, 2, 3, and 4 as inputs, form a pair of petals in each of the four quadrants of the polar coordinate system, and finally fuse the vibration information from each sensor. S4: Construct a hybrid receptive field residual convolutional neural network model, and use information fusion and multi-receptive field residual network structure to learn the fault characteristics of the synchronous condenser rotor; S5: Construct a multi-sensor fusion map dataset sample; S6: Data partitioning and model training; S7: Synchronous Condenser Rotor Fault Diagnosis. Input the vibration sequence to be tested and output the matrix sequence labels to be classified. The steps in step S4 for building the hybrid receptive field residual convolutional neural network model are as follows: First, a CBS module is built, and a batch normalization layer is added after the convolution operation to standardize the samples. The silu activation function is used to replace the convolution operation in the traditional convolutional neural network, so as to obtain a matrix of effective activation weights. Secondly, a fast spatial pyramid pooling (SPPF) module is constructed. The matrix extracted by the CBS module is subjected to pooling operations of 1×1, 3×3, and 5×5 subgraph partitioning, and the feature matrices of the three scales are connected to fuse the feature matrices of the three different scales. Finally, to address the shortcomings of traditional CNN structures and to incorporate the complex characteristics of camera fault information, a hybrid receptive field residual convolutional neural network structure, MRFRCNN, is proposed.
2. The method for diagnosing camera rotor faults based on SDP-MRFRCNN multi-receptive field information fusion convolutional neural network according to claim 1, characterized in that, The multi-sensor information fusion step in step S2 is as follows: First, the raw signals from each sensor in the three states of the camera are decomposed using SDP and displayed in the first quadrant. Secondly, the vibration characteristics of the four sensors are fused to form a pair of petals in each of the four quadrants of the polar coordinate system.
3. The method for diagnosing camera rotor faults based on SDP-MRFRCNN multi-receptive field information fusion convolutional neural network according to claim 2, characterized in that, During the SDP feature fusion process of the camera condenser vibration signal, the signals of sensors 1, 2, 3, and 4 were used as inputs, and the mirror symmetry was set to 45º, 135º, 225º, and 315º, with a parameter of 45°.
4. The method for diagnosing camera rotor faults based on a multi-receptive-field information fusion convolutional neural network using SDP-MRFRCNN according to claim 3, characterized in that, The structure of MRFRCNN: Its input layer is a 128x128 original image. After passing through the first 3x3 CBS layer, the data is split into three branches: Branch 1: 1x1 convolutional CBS branch; Branch 2: 3x3 CBS branch; Branch 3: 5x5 CBS branch. All three branches use two CBS layers with 16 and 32 kernels respectively, and a stride of 2. The outputs of branch 2 and branch 3 undergo an ADD operation, followed by a 1x1 point convolutional CBS layer to smooth the data flow. Then, a CONCAT operation is performed with the output of branch 1, forming a residual connection with branch 1. This reduces model complexity and prevents overfitting, as well as gradient vanishing or gradient exploding. After CONCAT of the three branches, the SPPF module performs dimensionality reduction, using multiple windows to extract spatial feature information of different sizes, improving the model's robustness to spatial layout and object deformation. Finally, two fully connected layers are used, and multi-class classification is achieved through Softmax.
5. The method for diagnosing camera rotor faults based on SDP-MRFRCNN multi-receptive field information fusion convolutional neural network according to claim 1, characterized in that, The multi-sensor fusion map dataset samples were constructed, trained, and divided into training and testing sets. The vibration SDP images of the camera were generated after information fusion, with 250 data points for each of the three states: normal, rubbing, and unbalanced. After mixing the data, random numbers were set to shuffle the data, and the training and testing sets were divided in a 4:1 ratio. The first three states were sequentially encoded as 0, 1, and 2 when input into the network.
6. The method for diagnosing camera rotor faults based on SDP-MRFRCNN multi-receptive field information fusion convolutional neural network according to claim 5, characterized in that, Based on the aforementioned dataset, in the PyTorch environment, the network learning rate was set to 0.0001, the batch size to 32, and the iterations to 50 rounds. The length of the input one-dimensional vibration sequence of the camera was L. in， The size of the two-dimensional image matrix fused by SDP multi-sensor feature fusion is a*b.