Search

CN-122023460-A - Three-dimensional vibration displacement measurement method of rotator based on deep learning and motion amplification by using binocular camera

CN122023460ACN 122023460 ACN122023460 ACN 122023460ACN-122023460-A

Abstract

The invention discloses a method for measuring three-dimensional vibration displacement of a rotating body based on deep learning and motion amplification by using a binocular camera, which comprises the steps of collecting a calibration image in a static state of the rotating body; the method comprises the steps of constructing a first data set under a rotating body motion state, carrying out vibration characteristic enhancement on a rotating body vibration image in the first data set according to a motion amplification algorithm based on a phase to obtain a second data set, constructing a third data set according to the second data set, training and verifying an improved YOLOv deep learning detection model by utilizing a training set and a verification set in the third data set to obtain a freezing model, carrying out rectangular frame detection on a test set in the third data set according to the freezing model to obtain a fourth data set, taking a vibration image of the test set and a rectangular detection frame in the fourth data set as input of a target tracking network, taking a result obtained through the target tracking network as a fifth data set, and obtaining three-dimensional vibration displacement data of the rotating body according to a calibration image and the fifth data set. According to the invention, sub-pixel level center point matching is realized through the improved YOLOv network, the motion amplification is combined to enhance the target tracking of the micro vibration signal characteristics, and the correlation of target tracking optimization time sequence data is introduced, so that the accuracy and the robustness of the three-dimensional vibration measurement method are effectively improved.

Inventors

  • WANG SEN
  • Lv juan
  • YANG FAMENG
  • ZHU HAILONG

Assignees

  • 昆明理工大学

Dates

Publication Date
20260512
Application Date
20260210

Claims (8)

  1. 1. A three-dimensional vibration displacement measuring method of a rotator based on deep learning and motion amplification by using a binocular camera is characterized by comprising the following steps: step 1, synchronously acquiring a plurality of calibration plate images at different positions and angles through two image acquisition devices with the same resolution ratio in a static state of a rotating body to serve as calibration images; Step 2, under the state of moving the rotating body, acquiring continuous rotating body vibration images by using two image acquisition devices with the same resolution, and splicing the rotating body vibration images of the same time frame synchronously acquired by the two image acquisition devices to construct a first data set, wherein the vertical distance between the optical centers of the first image acquisition device and the second image acquisition device and the center of the rotating body is the same, the first image acquisition device is positioned right in front of the rotating body, and the second image acquisition device and the first image acquisition device are arranged at intervals; Step 3, carrying out vibration characteristic enhancement on the rotating body vibration image in the first data set according to a motion amplification algorithm based on the phase to obtain a second data set; Step 4, randomly extracting a first preset number of vibration images from a second data set to carry out rectangular detection frame labeling, dividing the labeled vibration images into a training set and a verification set according to a preset proportion, extracting a continuous second preset number of vibration images from the second data set as a test set, and further forming a third data set consisting of the training set, the verification set and the test set, wherein the first preset number is smaller than the second preset number; Step 5, training and verifying the improved YOLOv deep learning detection model by using a training set and a verification set in the third data set to obtain a freezing model; Step 6, taking the vibration image of the test set in the third data set and the rectangular detection frame in the fourth data set as input of a target tracking network, and taking a result obtained through the target tracking network as a fifth data set; And 7, according to the acquired calibration image and the central point coordinate of the detection frame in the fifth data set, acquiring the relationship between the internal parameter and the external parameter of the binocular camera and the relative pose by using camera calibration, and calculating the three-dimensional space coordinate of the corresponding rotating body under the first image acquisition device by using a three-dimensional reconstruction principle, thereby acquiring three-dimensional vibration displacement data of the rotating body.
  2. 2. The method for measuring three-dimensional vibration displacement of a rotating body based on deep learning and motion amplification by using a binocular camera according to claim 1 is characterized in that a YOLOv is taken as a frame of an improved YOLOv-11 deep learning detection model, an SPPF module in YOLOv is replaced by a AIFI module so as to realize cross-layer attention dynamic fusion of multi-scale features, a C3K2 module in YOLOv is replaced by a CMSA module so as to construct multi-scale attention branches, and the capturing capability of rotating detail features is enhanced.
  3. 3. The method for measuring three-dimensional vibration displacement of a rotating body based on deep learning and motion amplification by using a binocular camera according to claim 2, wherein the AIFI module is specifically: simultaneously, based on the wide dimension and the high dimension of the 2D feature map, firstly generating grid coordinates in the wide direction and the high direction, then respectively calculating sine/cosine components in the wide direction and the high direction, then splicing, and finally obtaining a position code matched with the sequence length; The position codes are input to a multi-head self-attention module, attention characteristics output by the multi-head self-attention module are spliced with the original input three-dimensional characteristic diagram of the AIFI module, and normalization is completed through a Norm layer; and inputting the normalized characteristics into a feedforward neural network, splicing the feedforward characteristics output by the feedforward neural network and the normalized output of the multi-head self-attention module, and then finishing normalization through a Norm layer to restore into a three-dimensional characteristic diagram consistent with the input dimension.
  4. 4. The method for measuring three-dimensional vibration displacement of a rotating body based on deep learning and motion amplification by using a binocular camera according to claim 2, wherein the CMSA module is specifically: The method comprises the steps of taking a three-dimensional feature map as input, firstly carrying out preliminary mapping on input features through a first 1 multiplied by 1 convolution layer, and completing preliminary integration of features while keeping the number of channels unchanged; And splicing the MSA module output features and residual branches along the channel dimension, finishing channel information aggregation and dimension calibration of the features through a third 1X 1 convolution layer, and finally outputting a three-dimensional feature map consistent with the input dimension.
  5. 5. The method for measuring three-dimensional vibration displacement of a rotating body based on deep learning and motion amplification by using a binocular camera according to claim 4 is characterized in that an MSA module firstly splits input main branch characteristics into two branches along a channel dimension, namely a first branch 1 and a second branch 1, the first branch 1 is input into a 3×3 convolution layer to conduct basic local extraction on the characteristics, then splits output characteristics of the 3×3 convolution layer into two branches along the channel dimension, namely a first branch 2 and a second branch 2, the first branch 2 is input into a5×5 depth separable convolution layer to complete capturing of mesoscale space information, then splits output characteristics of the 5×5 depth separable convolution layer into two branches along the channel dimension, namely a first branch 3 and a second branch 3, the first branch 3 is input into a 7×7 depth separable convolution layer to further extract large-scale space characteristics, then inputs output results of the second branches 2, the second branches 3 and the 7×7 depth separable convolution layer after the output of the channel dimension is conducted into the second 1×1 convolution layer, and finally fuses the output characteristics of the second convolution layer 1 and the second branch 1.
  6. 6. The three-dimensional vibration displacement measurement method of a rotating body based on depth learning and motion amplification using a binocular camera according to claim 1, wherein three-dimensional space coordinates (X, Y, Z) of the rotating body are expressed as: ; Wherein, the In order to rotate the sub-matrix, 、 The coordinates of the center points of the detection frames of the corresponding images of the first image acquisition device and the second image acquisition device which are respectively output by the target tracking network, And Corresponding to the focal length parameters of the first image acquisition device and the second image acquisition device respectively, , Is a translation sub-vector.
  7. 7. A three-dimensional vibration displacement measuring system of a rotating body based on deep learning and motion amplification by using a binocular camera, characterized by comprising a module of the method of any one of claims 1-6.
  8. 8. A terminal device comprising a memory, a processor and a computer program stored on the memory, characterized in that the processor executes the computer program to carry out the steps of the method according to any one of claims 1-6.

Description

Three-dimensional vibration displacement measurement method of rotator based on deep learning and motion amplification by using binocular camera Technical Field The invention relates to a three-dimensional vibration displacement measurement method of a rotating body based on deep learning and motion amplification by using a binocular camera, belonging to the field of three-dimensional vibration displacement measurement based on computer vision. Background The running reliability of the rotary machine as a core power component of high-end equipment such as an aero-engine, a gas turbine and the like is directly related to the equipment performance and service life. Since continuous operation of rotary machines is prone to mechanical failure, even leading to equipment performance decay and even system failure, real-time monitoring of rotary machine displacement characteristics is critical to state assessment and life prediction. In the long-term service process of the rotary machine, the vibration of the rotary machine is often represented by high-frequency micro-amplitude characteristics of three-dimensional displacement and torsional coupling, two-dimensional measurement is difficult to accurately represent, and the amplitude is usually in the sub-millimeter level. In addition, due to the fact that the texture of the surface of the metal rotor is missing, characteristic instability and matching errors are easy to cause, so that the existing three-dimensional vibration measurement method has obvious technical blank in the aspects of precision and stability, and development of a novel visual measurement scheme is needed to solve the problems of capturing micro vibration and three-dimensional dynamic characterization. Disclosure of Invention The invention provides a three-dimensional vibration displacement measuring method of a rotating body based on deep learning and motion amplification by using a binocular camera, which realizes sub-pixel level center point matching by introducing a YOLOv network improved by a CMSA module and a AIFI module, enhances micro vibration signal characteristic target tracking by combining motion amplification, introduces target tracking optimization time sequence data relevance, and effectively improves the accuracy and the robustness of the three-dimensional vibration measuring method. The technical scheme of the invention is as follows: according to a first aspect of the present invention, there is provided a three-dimensional vibration displacement measurement method of a rotating body based on deep learning and motion amplification using a binocular camera, comprising: step 1, synchronously acquiring a plurality of calibration plate images at different positions and angles through two image acquisition devices with the same resolution ratio in a static state of a rotating body to serve as calibration images; Step 2, under the state of moving the rotating body, acquiring continuous rotating body vibration images by using two image acquisition devices with the same resolution, and splicing the rotating body vibration images of the same time frame synchronously acquired by the two image acquisition devices to construct a first data set, wherein the vertical distance between the optical centers of the first image acquisition device and the second image acquisition device and the center of the rotating body is the same, the first image acquisition device is positioned right in front of the rotating body, and the second image acquisition device and the first image acquisition device are arranged at intervals; Step 3, carrying out vibration characteristic enhancement on the rotating body vibration image in the first data set according to a motion amplification algorithm based on the phase to obtain a second data set; Step 4, randomly extracting a first preset number of vibration images from a second data set to carry out rectangular detection frame labeling, dividing the labeled vibration images into a training set and a verification set according to a preset proportion, extracting a continuous second preset number of vibration images from the second data set as a test set, and further forming a third data set consisting of the training set, the verification set and the test set, wherein the first preset number is smaller than the second preset number; Step 5, training and verifying the improved YOLOv deep learning detection model by using a training set and a verification set in the third data set to obtain a freezing model; Step 6, taking the vibration image of the test set in the third data set and the rectangular detection frame in the fourth data set as input of a target tracking network, and taking a result obtained through the target tracking network as a fifth data set; And 7, according to the acquired calibration image and the central point coordinate of the detection frame in the fifth data set, acquiring the relationship between the internal parameter and the external param